Initial commit: Nat20 Notes — TTRPG session transcription & summarization

2026-07-06 23:45:54 +12:00
commit 0f9e9d4c5c
49 changed files with 2926 additions and 0 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -0,0 +1,8 @@
+__pycache__/
+*.pyc
+.venv/
+node_modules/
+dist/
+*.db
+.env
+docker-compose.yml
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -0,0 +1,36 @@
+# Nat20 Notes — agent guide
+
+## Commands
+
+- `docker compose up -d --build` — build & launch both services (backend :8000, frontend :8020)
+- `cd frontend && npm run dev` — Vite dev server (proxies `/api` → `localhost:8000`)
+- **No tests, no lint/typecheck scripts exist.** Don't look for them.
+
+## Architecture
+
+- Two Docker Compose services: `backend` (Python 3.11, FastAPI, SQLite, WhisperX as lib) and `frontend` (React 18 + Vite + Tailwind, served by nginx which proxies `/api` → backend).
+- Data lives in Docker volume `app_data` at `/data` (SQLite DB + `uploads/audio/transcripts/notes` subdirs). No external DB.
+- Background jobs (transcribe, summarize) run via in-process `ThreadPoolExecutor(max_workers=2)` — no Redis/Celery.
+- Frontend serves on port **8020** (d20 nod).
+
+## Critical convention: flat settings schema
+
+Settings are a flat key/value table in SQLite. **Every layer must share exact same keys:**
+`backend/app/config.py:DEFAULT_SETTINGS`, `frontend/src/api.ts:AppSettings`, the settings router, and Setup/Settings pages. Changing a key without updating all layers silently desyncs the UI.
+
+## Docker / framework quirks
+
+- Torch pinned to `2.3.1` with `cu121` wheels (`--extra-index-url https://download.pytorch.org/whl/cu121`) — cuDNN ABI compat with `ctranslate2`. Don't bump unpinned.
+- Backend Dockerfile is two-stage (builder + runtime). ffmpeg required in runtime for audio extraction.
+- nginx `client_max_body_size` 500M, proxy timeouts 3600s. Files >15MB auto-chunked by frontend `createSession()`.
+- Frontend changes in Compose mode require a rebuild (`docker compose up -d --build`). Vite dev server is for local-only dev.
+
+## Tailwind theme
+
+Custom colors (`deep`, `panel`, `brass`, `ember`, etc.) and fonts (`Fraunces`, `Inter`, `IBM Plex Mono`) in `tailwind.config.js`. Reusable component classes in `frontend/src/styles.css` (`.card`, `.btn-primary`, `.btn-secondary`, `.input`, `.eyebrow`) — use these over raw Tailwind utilities.
+
+## Constraints
+
+- Requires NVIDIA Container Toolkit + HuggingFace token for speaker diarization.
+- First run: setup wizard, then `onboarding_completed` flag gates access to main UI.
+- Default Ollama endpoint: `http://localhost:11434`.
--- a/21
+++ b/21
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Nat20 Notes
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
--- a/README.md
+++ b/README.md
@@ -0,0 +1,91 @@
+# Nat20 Notes
+
+Turn a recorded tabletop RPG session (audio or video) into two documents:
+a full GM/DM session log, and a spoiler-free player recap — using local
+transcription (WhisperX) and either a local LLM (Ollama) or any
+OpenAI-compatible hosted API for summarization.
+
+## Requirements
+
+- Docker + Docker Compose
+- An NVIDIA GPU with drivers + [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) installed on the host (for transcription; CPU-only works but is much slower)
+- A free [HuggingFace token](https://huggingface.co/settings/tokens) — needed for speaker diarization. You'll also need to accept the terms on the gated pyannote model page it links you to on first run.
+- Either: [Ollama](https://ollama.com) running somewhere reachable from this app (local or LAN), **or** an API key for a hosted LLM (OpenAI, or any OpenAI-compatible provider)
+
+## Quick start
+
+```bash
+git clone <this-repo>
+cd nat20-notes
+# Review docker-compose.yml — see Configuration below for env vars
+docker compose up -d --build
+```
+
+Then open `http://<your-server>:8020` (yes, that port's a nod to the d20) and follow the setup wizard:
+
+1. Choose a Whisper model size based on your GPU's available VRAM (guidance shown in-app)
+2. Choose local Ollama or a hosted API for summarization, and paste your HF token
+3. Optionally paste campaign/world context (NPC names, places) so summaries recognize them correctly
+
+A reference compose file using named volumes only (no host paths) is at
+[`docker-compose.example.yml`](./docker-compose.example.yml).
+
+## Using it
+
+1. **Upload** a recording (audio or video — video is auto-converted, audio-only files skip that step and are much faster). Files over **90 MB** are auto-chunked with 3-way concurrent uploads.
+2. **Transcribe** — runs in the background. On completion you're **automatically taken** to the speaker-naming screen.
+3. **Name your speakers** — a waveform-style "session reel" shows each detected speaker's segments; click any point to jump the audio there and hear who's talking, then type in their name. Use the checkboxes to **merge** speakers (e.g. when diarization over-splits one person into `SPEAKER_00`, `SPEAKER_05`, etc.).
+4. **Generate notes** — click "Done naming" and confirm; notes generation starts automatically and you're taken to the job progress screen. When finished, navigate to the notes viewer.
+5. **Review & tweak** — regenerate notes, rename speakers, or delete jobs/sessions from the session detail page.
+
+## Data layout
+
+The app stores everything under `/data` (inside the container), which by default
+is a [bind mount](./docker-compose.yml) to a host path of your choice:
+
+| Directory / File | Contents |
+|---|---|
+| `audio/` | Uploaded recordings and extracted audio |
+| `transcriptions/` | Per-session transcript JSON files |
+| `notes/` | Generated notes (GM log + player recap) |
+| `app.db` | SQLite database (sessions, speakers, settings, jobs) |
+
+## Configuration
+
+### Prefilling the setup wizard
+
+Set `NAT20_*` environment variables under the `backend` service in
+`docker-compose.yml` — the wizard will pick them up as defaults:
+
+```yaml
+environment:
+  NAT20_HF_TOKEN: "hf_..."
+  NAT20_WHISPER_MODEL: medium
+  NAT20_OLLAMA_HOST: http://192.168.0.16:11434
+  NAT20_WORLD_CONTEXT_PATH: /data/campaign-context.txt
+```
+
+See the `environment:` block in `docker-compose.yml` for the full list.
+
+### Campaign context
+
+You can paste context directly in the Settings page, or point to a file
+inside the container using the `world_context_path` setting (or the
+`NAT20_WORLD_CONTEXT_PATH` env var). The file path version is useful for
+large campaign bibles that you update independently.
+
+## Notes on hardware
+
+Transcription is GPU-bound and by far the slowest step for long sessions.
+Summarization is comparatively light — a 7-8B parameter local model is
+sufficient for most groups; only step up in size if you have the VRAM
+headroom after Whisper's footprint (they don't run at the same time, so
+you only need enough VRAM for whichever is currently running, plus normal
+system overhead from other GPU-using services).
+
+## Architecture
+
+- `backend/` — FastAPI, SQLite (no external DB needed), WhisperX as a library (no nested Docker), in-process `ThreadPoolExecutor` background jobs (no Redis/Celery)
+- `frontend/` — React 18 + Vite + Tailwind, served via nginx (listens on port **8020**) which proxies `/api` to the backend
+
+Both run as standard Docker Compose services — no special orchestration needed beyond GPU passthrough for the backend.
--- a/backend/.dockerignore
+++ b/backend/.dockerignore
@@ -0,0 +1,10 @@
+__pycache__/
+*.pyc
+*.pyo
+.pytest_cache/
+.mypy_cache/
+.venv/
+venv/
+*.egg-info/
+.git/
+.env
--- a/backend/Dockerfile
+++ b/backend/Dockerfile
@@ -0,0 +1,52 @@
+# ==========================================
+# STAGE 1: Builder Environment
+# ==========================================
+FROM python:3.11-slim AS builder
+
+WORKDIR /build
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential \
+    gcc \
+    patchelf \
+    && rm -rf /var/lib/apt/lists/*
+
+ENV PIP_NO_CACHE_DIR=1
+ENV PIP_DISABLE_PIP_VERSION_CHECK=1
+
+COPY requirements.txt .
+
+RUN pip install --user --extra-index-url https://download.pytorch.org/whl/cu121 -r requirements.txt
+
+# Silence the "Lightning auto-upgraded checkpoint" warning on every diarization run
+RUN python -c "import sys, subprocess, whisperx, pathlib; ckpt = pathlib.Path(whisperx.__file__).parent/'assets'/'pytorch_model.bin'; ckpt.exists() and subprocess.run([sys.executable, '-m', 'pytorch_lightning.utilities.upgrade_checkpoint', str(ckpt)])" 2>/dev/null || true
+
+RUN patchelf --clear-execstack /root/.local/lib/python3.11/site-packages/ctranslate2.libs/libctranslate2-*.so*
+
+RUN find /root/.local -type d -name "__pycache__" -exec rm -rf {} + \
+    && find /root/.local -type d \( -name "test" -o -name "tests" \) -exec rm -rf {} + \
+    && find /root/.local -name "*.dist-info" -exec sh -c 'rm -f "$1"/RECORD "$1"/INSTALLER' _ {} \;
+
+# ==========================================
+# STAGE 2: Lightweight Final Runtime
+# ==========================================
+FROM python:3.11-slim AS runtime
+
+WORKDIR /app
+
+# ffmpeg is mandatory for the audio-extraction step.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ffmpeg \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copy the (now stripped) site-packages from the builder stage. None of the
+# build-essential/gcc toolchain from Stage 1 ends up here.
+COPY --from=builder /root/.local /root/.local
+COPY app /app/app
+
+ENV PATH=/root/.local/bin:$PATH
+ENV PYTHONUNBUFFERED=1
+
+EXPOSE 8000
+
+CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000"]
--- a/backend/app/init.py
+++ b/backend/app/init.py
--- a/backend/app/config.py
+++ b/backend/app/config.py
@@ -0,0 +1,49 @@
+"""
+App-wide configuration. Everything is either an environment variable (set once,
+at deploy time) or a row in the `settings` table (editable at runtime via the
+setup wizard / settings page) - never a hardcoded path or host.
+"""
+import os
+from pathlib import Path
+
+# Deploy-time config (env vars) - where things live on disk inside the container.
+DATA_DIR = Path(os.environ.get("APP_DATA_DIR", "/data"))
+UPLOAD_DIR = DATA_DIR / "audio"
+AUDIO_DIR = DATA_DIR / "audio"
+TRANSCRIPT_DIR = DATA_DIR / "transcriptions"
+NOTES_DIR = DATA_DIR / "notes"
+DB_PATH = DATA_DIR / "app.db"
+
+for d in (UPLOAD_DIR, AUDIO_DIR, TRANSCRIPT_DIR, NOTES_DIR):
+    d.mkdir(parents=True, exist_ok=True)
+
+# Runtime-editable settings (stored in DB, these are just first-run defaults).
+# NOTE: this flat key/value schema is the single source of truth for settings -
+# it's what every consumer (pipeline/summarize.py, pipeline/transcribe.py,
+# routers/models.py, routers/diagnostics.py) actually reads. The settings
+# router and frontend must mirror these exact keys - a nested schema would
+# silently disconnect the Settings/Setup UI from the pipeline.
+DEFAULT_SETTINGS = {
+    # Onboarding
+    "onboarding_completed": "false",
+
+    # Transcription
+    "whisper_model": "medium",       # tiny|base|small|medium|large-v3 - user picks based on their hardware
+    "whisper_compute_type": "int8",
+    "hf_token": "",                  # required for diarization (pyannote gated models)
+
+    # Summarization backend: "ollama" (local) or "api" (hosted, OpenAI-compatible)
+    "llm_mode": "ollama",
+    "ollama_host": "http://localhost:11434",
+    "ollama_model": "qwen2.5:7b",
+    "api_base_url": "https://api.openai.com/v1",
+    "api_key": "",
+    "api_model": "gpt-4o-mini",
+
+    # Chunking for long transcripts (map-reduce summarization)
+    "chunk_word_target": "2500",
+
+    # Optional world context injected into every summarization prompt
+    "world_context": "",
+    "world_context_path": "",
+}
--- a/backend/app/database.py
+++ b/backend/app/database.py
@@ -0,0 +1,136 @@
+import json
+import sqlite3
+import threading
+import time
+import uuid
+from contextlib import contextmanager
+
+from . import config
+
+_local = threading.local()
+
+
+def get_conn():
+    if not hasattr(_local, "conn"):
+        _local.conn = sqlite3.connect(config.DB_PATH, check_same_thread=False)
+        _local.conn.row_factory = sqlite3.Row
+        _local.conn.execute("PRAGMA foreign_keys = ON")
+    return _local.conn
+
+
+@contextmanager
+def tx():
+    conn = get_conn()
+    try:
+        yield conn
+        conn.commit()
+    except Exception:
+        conn.rollback()
+        raise
+
+
+def init_db():
+    with tx() as conn:
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS settings (
+                key TEXT PRIMARY KEY,
+                value TEXT NOT NULL
+            )
+        """)
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS sessions (
+                id TEXT PRIMARY KEY,
+                name TEXT NOT NULL,
+                original_filename TEXT NOT NULL,
+                video_path TEXT,
+                audio_path TEXT,
+                transcript_path TEXT,
+                status TEXT NOT NULL DEFAULT 'uploaded',
+                created_at REAL NOT NULL
+            )
+        """)
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS speakers (
+                id TEXT PRIMARY KEY,
+                session_id TEXT NOT NULL REFERENCES sessions(id) ON DELETE CASCADE,
+                raw_label TEXT NOT NULL,
+                display_name TEXT,
+                UNIQUE(session_id, raw_label)
+            )
+        """)
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS notes (
+                session_id TEXT PRIMARY KEY REFERENCES sessions(id) ON DELETE CASCADE,
+                dm_notes TEXT,
+                player_recap TEXT,
+                generated_at REAL
+            )
+        """)
+        conn.execute("""
+            CREATE TABLE IF NOT EXISTS jobs (
+                id TEXT PRIMARY KEY,
+                session_id TEXT NOT NULL REFERENCES sessions(id) ON DELETE CASCADE,
+                job_type TEXT NOT NULL,        -- 'transcribe' | 'summarize'
+                status TEXT NOT NULL DEFAULT 'queued',  -- queued|running|done|error
+                progress TEXT,                 -- free-text progress message
+                error TEXT,
+                error_stage TEXT,               -- which pipeline stage failed, e.g. 'transcription'
+                error_detail TEXT,               -- full traceback, for logs/advanced view only
+                created_at REAL NOT NULL,
+                updated_at REAL NOT NULL
+            )
+        """)
+        # seed defaults if empty
+        existing = {r["key"] for r in conn.execute("SELECT key FROM settings")}
+        for k, v in config.DEFAULT_SETTINGS.items():
+            if k not in existing:
+                conn.execute("INSERT INTO settings (key, value) VALUES (?, ?)", (k, v))
+
+
+def get_settings() -> dict:
+    conn = get_conn()
+    rows = conn.execute("SELECT key, value FROM settings").fetchall()
+    return {r["key"]: r["value"] for r in rows}
+
+
+def update_settings(patch: dict):
+    with tx() as conn:
+        for k, v in patch.items():
+            conn.execute(
+                "INSERT INTO settings (key, value) VALUES (?, ?) "
+                "ON CONFLICT(key) DO UPDATE SET value=excluded.value",
+                (k, str(v)),
+            )
+
+
+def new_id() -> str:
+    return uuid.uuid4().hex[:12]
+
+
+def now() -> float:
+    return time.time()
+
+
+def create_job(session_id: str, job_type: str) -> str:
+    job_id = new_id()
+    with tx() as conn:
+        conn.execute(
+            "INSERT INTO jobs (id, session_id, job_type, status, created_at, updated_at) "
+            "VALUES (?, ?, ?, 'queued', ?, ?)",
+            (job_id, session_id, job_type, now(), now()),
+        )
+    return job_id
+
+
+def update_job(job_id: str, **fields):
+    if not fields:
+        return
+    fields["updated_at"] = now()
+    cols = ", ".join(f"{k} = ?" for k in fields)
+    with tx() as conn:
+        conn.execute(f"UPDATE jobs SET {cols} WHERE id = ?", (*fields.values(), job_id))
+
+
+def get_job(job_id: str) -> dict | None:
+    row = get_conn().execute("SELECT * FROM jobs WHERE id = ?", (job_id,)).fetchone()
+    return dict(row) if row else None
--- a/backend/app/diagnostics.py
+++ b/backend/app/diagnostics.py
@@ -0,0 +1,73 @@
+import shutil
+import subprocess
+
+import requests
+
+from . import database as db, config
+
+
+def _check_ffmpeg() -> dict:
+    path = shutil.which("ffmpeg")
+    if not path:
+        return {"name": "ffmpeg", "ok": False, "message": "ffmpeg not found on PATH. This should be baked into the Docker image - if you're seeing this, the image build is broken."}
+    return {"name": "ffmpeg", "ok": True, "message": path}
+
+
+def _check_gpu() -> dict:
+    try:
+        import torch
+        if torch.cuda.is_available():
+            name = torch.cuda.get_device_name(0)
+            free, total = torch.cuda.mem_get_info(0)
+            free_gb, total_gb = free / 1e9, total / 1e9
+            return {"name": "gpu", "ok": True, "message": f"{name} — {free_gb:.1f}GB free of {total_gb:.1f}GB"}
+        return {"name": "gpu", "ok": False, "message": "No CUDA GPU visible. Check that the container was started with --gpus all / GPU passthrough in compose, and that host NVIDIA drivers + Container Toolkit are installed. Transcription will fall back to CPU, which is much slower."}
+    except Exception as e:
+        return {"name": "gpu", "ok": False, "message": f"Could not query GPU: {e}"}
+
+
+def _check_hf_token() -> dict:
+    settings = db.get_settings()
+    token = settings.get("hf_token", "")
+    if not token:
+        return {"name": "huggingface_token", "ok": False, "message": "No HF token set. Diarization (speaker separation) will fail without one. Add it in Settings."}
+    return {"name": "huggingface_token", "ok": True, "message": "Token is set (not validated against HF until first diarization run)."}
+
+
+def _check_llm_backend() -> dict:
+    settings = db.get_settings()
+    if settings.get("llm_mode") == "api":
+        if not settings.get("api_key"):
+            return {"name": "llm_backend", "ok": False, "message": "Hosted API selected but no API key set."}
+        return {"name": "llm_backend", "ok": True, "message": f"Hosted API configured ({settings.get('api_base_url')})"}
+
+    host = settings.get("ollama_host", "http://localhost:11434")
+    model = settings.get("ollama_model", "")
+    try:
+        resp = requests.get(f"{host.rstrip('/')}/api/tags", timeout=5)
+        resp.raise_for_status()
+        models = [m["name"] for m in resp.json().get("models", [])]
+        if model not in models:
+            return {"name": "llm_backend", "ok": False, "message": f"Reached Ollama at {host}, but model '{model}' isn't pulled there. Available: {', '.join(models) or '(none)'}. Run: ollama pull {model}"}
+        return {"name": "llm_backend", "ok": True, "message": f"Ollama reachable at {host}, model '{model}' available"}
+    except requests.exceptions.ConnectionError:
+        return {"name": "llm_backend", "ok": False, "message": f"Could not connect to Ollama at {host}. Check the host/port in Settings, and that Ollama is running and reachable from this container."}
+    except Exception as e:
+        return {"name": "llm_backend", "ok": False, "message": f"Error checking Ollama: {e}"}
+
+
+def _check_disk_space() -> dict:
+    usage = shutil.disk_usage(config.DATA_DIR)
+    free_gb = usage.free / 1e9
+    ok = free_gb > 5
+    return {"name": "disk_space", "ok": ok, "message": f"{free_gb:.1f}GB free in {config.DATA_DIR}" + ("" if ok else " — this is low, recordings and model caches need room")}
+
+
+def run_diagnostics() -> list[dict]:
+    return [
+        _check_ffmpeg(),
+        _check_gpu(),
+        _check_hf_token(),
+        _check_llm_backend(),
+        _check_disk_space(),
+    ]
--- a/backend/app/errors.py
+++ b/backend/app/errors.py
@@ -0,0 +1,29 @@
+class PipelineError(Exception):
+    """Base class for pipeline failures. `stage` and `message` are shown to the
+    user directly; the original exception (if any) is kept for logs only."""
+
+    def __init__(self, stage: str, message: str, cause: Exception | None = None):
+        self.stage = stage
+        self.message = message
+        self.cause = cause
+        super().__init__(f"[{stage}] {message}")
+
+
+class AudioExtractionError(PipelineError):
+    def __init__(self, message: str, cause: Exception | None = None):
+        super().__init__("audio_extraction", message, cause)
+
+
+class TranscriptionError(PipelineError):
+    def __init__(self, message: str, cause: Exception | None = None):
+        super().__init__("transcription", message, cause)
+
+
+class DiarizationError(PipelineError):
+    def __init__(self, message: str, cause: Exception | None = None):
+        super().__init__("diarization", message, cause)
+
+
+class SummarizationError(PipelineError):
+    def __init__(self, message: str, cause: Exception | None = None):
+        super().__init__("summarization", message, cause)
--- a/backend/app/jobs.py
+++ b/backend/app/jobs.py
@@ -0,0 +1,41 @@
+"""
+Minimal background job runner.
+"""
+import traceback
+from concurrent.futures import ThreadPoolExecutor
+
+from . import database as db
+from .errors import PipelineError
+from .logging_config import get_logger
+
+log = get_logger(__name__)
+_executor = ThreadPoolExecutor(max_workers=2)  # transcription is GPU-bound anyway; no benefit to more workers
+
+
+def submit(job_id: str, fn, *args, **kwargs):
+    def _run():
+        log.info("Job %s starting", job_id)
+        db.update_job(job_id, status="running", progress="Starting...")
+        try:
+            def progress_cb(msg):
+                log.info("Job %s: %s", job_id, msg)
+                db.update_job(job_id, progress=msg)
+            fn(*args, progress_cb=progress_cb, **kwargs)
+            db.update_job(job_id, status="done", progress="Complete")
+            log.info("Job %s complete", job_id)
+        except PipelineError as e:
+            log.error("Job %s failed at stage '%s': %s", job_id, e.stage, e.message)
+            db.update_job(
+                job_id, status="error",
+                error=e.message, error_stage=e.stage,
+                error_detail=traceback.format_exc(),
+            )
+        except Exception as e:
+            log.exception("Job %s failed with unexpected error", job_id)
+            db.update_job(
+                job_id, status="error",
+                error=f"Unexpected error: {e}", error_stage="unknown",
+                error_detail=traceback.format_exc(),
+            )
+
+    _executor.submit(_run)
--- a/backend/app/logging_config.py
+++ b/backend/app/logging_config.py
@@ -0,0 +1,18 @@
+import logging
+import sys
+
+
+def setup_logging():
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s %(levelname)-8s %(name)s: %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+        stream=sys.stdout,
+    )
+    # quiet down noisy libraries so pipeline logs are easy to spot
+    for noisy in ("httpx", "urllib3", "pyannote"):
+        logging.getLogger(noisy).setLevel(logging.WARNING)
+
+
+def get_logger(name: str) -> logging.Logger:
+    return logging.getLogger(name)
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -0,0 +1,67 @@
+from fastapi import FastAPI, Request
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import FileResponse, JSONResponse
+from fastapi import HTTPException
+
+from . import database as db, config
+from .logging_config import setup_logging, get_logger
+from .errors import PipelineError
+from .routers import settings as settings_router
+from .routers import sessions as sessions_router
+from .routers import speakers as speakers_router
+from .routers import notes as notes_router
+from .routers import jobs as jobs_router
+from .routers import models as models_router
+from .routers import diagnostics as diagnostics_router
+
+setup_logging()
+log = get_logger(__name__)
+
+app = FastAPI(title="Nat20 Notes")
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+
+@app.exception_handler(PipelineError)
+async def pipeline_error_handler(request: Request, exc: PipelineError):
+    log.error("Pipeline error in %s: %s", exc.stage, exc.message, exc_info=exc.cause or exc)
+    return JSONResponse(status_code=502, content={"stage": exc.stage, "message": exc.message})
+
+
+@app.exception_handler(Exception)
+async def unhandled_error_handler(request: Request, exc: Exception):
+    log.exception("Unhandled error on %s %s", request.method, request.url.path)
+    return JSONResponse(status_code=500, content={"stage": "unknown", "message": f"Unexpected error: {exc}"})
+
+
+@app.on_event("startup")
+def on_startup():
+    db.init_db()
+    log.info("Database initialized at %s", config.DB_PATH)
+
+
+app.include_router(settings_router.router)
+app.include_router(sessions_router.router)
+app.include_router(speakers_router.router)
+app.include_router(notes_router.router)
+app.include_router(jobs_router.router)
+app.include_router(models_router.router)
+app.include_router(diagnostics_router.router)
+
+
+@app.get("/api/sessions/{session_id}/audio")
+def get_audio(session_id: str):
+    audio_path = config.AUDIO_DIR / f"{session_id}.wav"
+    if not audio_path.exists():
+        raise HTTPException(404, "Audio not available for this session")
+    return FileResponse(audio_path, media_type="audio/wav")
+
+
+@app.get("/api/health")
+def health():
+    return {"ok": True}
--- a/backend/app/pipeline/init.py
+++ b/backend/app/pipeline/init.py
--- a/backend/app/pipeline/audio.py
+++ b/backend/app/pipeline/audio.py
@@ -0,0 +1,35 @@
+import subprocess
+from pathlib import Path
+
+from ..errors import AudioExtractionError
+from ..logging_config import get_logger
+
+log = get_logger(__name__)
+
+
+def extract_audio(video_path: Path, audio_path: Path) -> Path:
+    """Extract 16kHz mono wav from any video/audio container."""
+    if not video_path.exists():
+        raise AudioExtractionError(f"Uploaded file not found on disk at {video_path}. It may not have finished uploading, or the upload volume isn't mounted correctly.")
+
+    if audio_path.exists():
+        log.info("Audio already extracted at %s, skipping", audio_path)
+        return audio_path
+
+    log.info("Extracting audio: %s -> %s", video_path, audio_path)
+    result = subprocess.run(
+        [
+            "ffmpeg", "-y", "-i", str(video_path),
+            "-vn", "-acodec", "pcm_s16le", "-ar", "16000", "-ac", "1",
+            str(audio_path),
+        ],
+        capture_output=True, text=True,
+    )
+    if result.returncode != 0:
+        # surface ffmpeg's own error line rather than the whole verbose log
+        stderr_tail = "\n".join(result.stderr.strip().splitlines()[-5:])
+        raise AudioExtractionError(
+            f"ffmpeg failed to extract audio (exit code {result.returncode}). "
+            f"This usually means the file is corrupt or not a supported format. Last output:\n{stderr_tail}"
+        )
+    return audio_path
--- a/backend/app/pipeline/summarize.py
+++ b/backend/app/pipeline/summarize.py
@@ -0,0 +1,152 @@
+"""
+Map-reduce summarization: chunk transcript -> per-chunk DM + player summaries
+-> combine into final notes. Backend-agnostic: works against a local Ollama
+instance or any OpenAI-compatible hosted API, selected via settings.
+"""
+from pathlib import Path
+
+import requests
+
+from .turns import chunk_turns, turns_to_text, fmt_time
+from ..errors import SummarizationError
+from ..logging_config import get_logger
+
+log = get_logger(__name__)
+
+DM_CHUNK_PROMPT = """You are summarizing a chunk of a tabletop RPG session transcript for the Game Master's private notes.
+{world_context_block}
+Extract concise bullet points:
+- Major plot events and decisions made
+- NPC interactions (names, what was said/promised/revealed)
+- Combat outcomes (who fought what, notable rolls, deaths/near-deaths)
+- Loot, rewards, or resources gained/lost
+- Any GM secrets, foreshadowing, or plot threads revealed at the table
+- Open questions or hooks left dangling
+
+Be concise and factual. Use the speaker names given. Do not invent anything not in the transcript.
+
+Transcript chunk (time range {time_range}):
+{chunk_text}
+"""
+
+PLAYER_CHUNK_PROMPT = """You are summarizing a chunk of a tabletop RPG session transcript for a player-facing recap.
+{world_context_block}
+Extract concise bullet points of what happened IN THE STORY from the players' in-character perspective:
+- Where the party went, who they met, what was said in-character
+- Combat encounters and outcomes
+- Loot/items the party found
+- Decisions the party made and their in-fiction consequences
+
+Exclude out-of-character rules discussion, GM asides, and anything that reads as a secret not yet revealed to characters.
+If unsure whether something is a spoiler, leave it out.
+
+Transcript chunk (time range {time_range}):
+{chunk_text}
+"""
+
+DM_FINAL_PROMPT = """Combine these chunk summaries from a single session into one cohesive GM session log.
+Organize into sections: Recap, NPC Interactions, Combat & Encounters, Loot & Rewards, Decisions & Consequences, Open Threads / Hooks for Next Session.
+Write in clear prose/bullets, no fluff, no repetition across chunks.
+
+Chunk summaries:
+{summaries}
+"""
+
+PLAYER_FINAL_PROMPT = """Combine these chunk summaries into a single "previously on" style recap for the players, in-character perspective only.
+Write it like a story recap they could reread before the next session. No GM secrets, no OOC content. Keep it engaging but concise.
+
+Chunk summaries:
+{summaries}
+"""
+
+
+def _call_ollama(host: str, model: str, prompt: str) -> str:
+    url = f"{host.rstrip('/')}/api/generate"
+    try:
+        resp = requests.post(
+            url,
+            json={"model": model, "prompt": prompt, "stream": False, "options": {"num_ctx": 8192}},
+            timeout=600,
+        )
+    except requests.exceptions.ConnectionError as e:
+        raise SummarizationError(
+            f"Could not connect to Ollama at {host}. Check the host/port in Settings and that Ollama is running.", cause=e
+        )
+    except requests.exceptions.Timeout as e:
+        raise SummarizationError(f"Ollama at {host} didn't respond within 10 minutes - it may be overloaded or stuck.", cause=e)
+
+    if resp.status_code == 404:
+        raise SummarizationError(
+            f"Ollama returned 404 for model '{model}'. This almost always means the model isn't pulled on that host. "
+            f"Run: ollama pull {model}", cause=None
+        )
+    try:
+        resp.raise_for_status()
+    except requests.exceptions.HTTPError as e:
+        raise SummarizationError(f"Ollama returned an error ({resp.status_code}): {resp.text[:300]}", cause=e)
+
+    return resp.json()["response"].strip()
+
+
+def _call_api(base_url: str, api_key: str, model: str, prompt: str) -> str:
+    url = f"{base_url.rstrip('/')}/chat/completions"
+    try:
+        resp = requests.post(
+            url,
+            headers={"Authorization": f"Bearer {api_key}"},
+            json={"model": model, "messages": [{"role": "user", "content": prompt}]},
+            timeout=600,
+        )
+    except requests.exceptions.ConnectionError as e:
+        raise SummarizationError(f"Could not connect to API at {base_url}. Check the base URL in Settings.", cause=e)
+    except requests.exceptions.Timeout as e:
+        raise SummarizationError(f"API at {base_url} didn't respond within 10 minutes.", cause=e)
+
+    if resp.status_code == 401:
+        raise SummarizationError("API rejected the request as unauthorized (401) - check your API key in Settings.", cause=None)
+    if resp.status_code == 404:
+        raise SummarizationError(f"API returned 404 - check the base URL and that model '{model}' exists for this provider.", cause=None)
+    try:
+        resp.raise_for_status()
+    except requests.exceptions.HTTPError as e:
+        raise SummarizationError(f"API returned an error ({resp.status_code}): {resp.text[:300]}", cause=e)
+
+    try:
+        return resp.json()["choices"][0]["message"]["content"].strip()
+    except (KeyError, IndexError) as e:
+        raise SummarizationError(f"API response didn't have the expected shape: {resp.text[:300]}", cause=e)
+
+
+def make_llm_caller(settings: dict):
+    if settings.get("llm_mode") == "api":
+        base_url, api_key, model = settings["api_base_url"], settings["api_key"], settings["api_model"]
+        return lambda prompt: _call_api(base_url, api_key, model, prompt)
+    host, model = settings["ollama_host"], settings["ollama_model"]
+    return lambda prompt: _call_ollama(host, model, prompt)
+
+
+def summarize_session(turns: list[dict], settings: dict, progress_cb=None) -> tuple[str, str]:
+    call = make_llm_caller(settings)
+    ctx_path = settings.get("world_context_path") or ""
+    if ctx_path and Path(ctx_path).exists():
+        world_context = Path(ctx_path).read_text(encoding="utf-8").strip()
+    else:
+        world_context = (settings.get("world_context") or "").strip()
+    world_block = f"\nCampaign context (use this to recognize names/places correctly):\n{world_context}\n" if world_context else ""
+
+    chunks = chunk_turns(turns, int(settings.get("chunk_word_target", 2500)))
+    dm_summaries, player_summaries = [], []
+
+    for i, chunk in enumerate(chunks):
+        if progress_cb:
+            progress_cb(f"Summarizing chunk {i+1}/{len(chunks)}")
+        time_range = f"{fmt_time(chunk[0]['start'])}-{fmt_time(chunk[-1]['end'])}"
+        chunk_text = turns_to_text(chunk)
+        dm_summaries.append(call(DM_CHUNK_PROMPT.format(world_context_block=world_block, time_range=time_range, chunk_text=chunk_text)))
+        player_summaries.append(call(PLAYER_CHUNK_PROMPT.format(world_context_block=world_block, time_range=time_range, chunk_text=chunk_text)))
+
+    if progress_cb:
+        progress_cb("Writing final notes...")
+    dm_final = call(DM_FINAL_PROMPT.format(summaries="\n\n".join(dm_summaries)))
+    player_final = call(PLAYER_FINAL_PROMPT.format(summaries="\n\n".join(player_summaries)))
+    return dm_final, player_final
--- a/backend/app/pipeline/transcribe.py
+++ b/backend/app/pipeline/transcribe.py
@@ -0,0 +1,104 @@
+"""
+Transcription + diarization using WhisperX as a Python library.
+"""
+import gc
+import json
+from pathlib import Path
+
+import torch
+import whisperx
+
+from ..errors import TranscriptionError, DiarizationError
+from ..logging_config import get_logger
+
+log = get_logger(__name__)
+
+
+def transcribe_and_diarize(
+    audio_path: Path,
+    output_json: Path,
+    model_size: str,
+    compute_type: str,
+    hf_token: str,
+    progress_cb=None,
+) -> dict:
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    if device == "cpu":
+        log.warning("No CUDA GPU available - running on CPU, this will be much slower")
+
+    def report(msg):
+        log.info(msg)
+        if progress_cb:
+            progress_cb(msg)
+
+    report("Loading transcription model...")
+    try:
+        model = whisperx.load_model(model_size, device, compute_type=compute_type)
+    except RuntimeError as e:
+        if "out of memory" in str(e).lower():
+            raise TranscriptionError(
+                f"Ran out of GPU memory loading the '{model_size}' model. "
+                f"Try a smaller model size in Settings (e.g. 'small' or 'medium'), "
+                f"or check nothing else is using the GPU right now.", cause=e
+            )
+        raise TranscriptionError(f"Failed to load Whisper model '{model_size}': {e}", cause=e)
+    except Exception as e:
+        raise TranscriptionError(f"Failed to load Whisper model '{model_size}': {e}", cause=e)
+
+    try:
+        report("Loading audio...")
+        audio = whisperx.load_audio(str(audio_path))
+
+        report("Transcribing...")
+        result = model.transcribe(audio, batch_size=16)
+    except RuntimeError as e:
+        if "out of memory" in str(e).lower():
+            raise TranscriptionError(
+                f"Ran out of GPU memory during transcription. Try a smaller Whisper model size in Settings.", cause=e
+            )
+        raise TranscriptionError(f"Transcription failed: {e}", cause=e)
+    except Exception as e:
+        raise TranscriptionError(f"Transcription failed: {e}", cause=e)
+    finally:
+        del model
+        gc.collect()
+        if device == "cuda":
+            torch.cuda.empty_cache()
+
+    try:
+        report("Aligning...")
+        align_model, metadata = whisperx.load_align_model(language_code=result["language"], device=device)
+        result = whisperx.align(result["segments"], align_model, metadata, audio, device, return_char_alignments=False)
+        del align_model
+        gc.collect()
+        if device == "cuda":
+            torch.cuda.empty_cache()
+    except Exception as e:
+        raise TranscriptionError(f"Alignment step failed: {e}", cause=e)
+
+    if hf_token:
+        try:
+            report("Diarizing...")
+            diarize_model = whisperx.diarize.DiarizationPipeline(use_auth_token=hf_token, device=device)
+            diarize_segments = diarize_model(audio)
+            result = whisperx.assign_word_speakers(diarize_segments, result)
+        except Exception as e:
+            msg = str(e)
+            if "gated" in msg.lower() or "403" in msg or "access" in msg.lower():
+                raise DiarizationError(
+                    "HuggingFace rejected access to the diarization model. Make sure you've accepted the "
+                    "terms on the gated model page (linked in the error below) using the SAME account "
+                    f"your HF token belongs to.\n{msg}", cause=e
+                )
+            raise DiarizationError(f"Diarization failed: {msg}", cause=e)
+    else:
+        report("No HF token set - skipping speaker diarization (all speech will be unattributed)")
+
+    try:
+        output_json.parent.mkdir(parents=True, exist_ok=True)
+        output_json.write_text(json.dumps(result, ensure_ascii=False), encoding='utf-8')
+    except OSError as e:
+        raise TranscriptionError(f"Transcribed successfully, but failed to write output to {output_json}: {e}", cause=e)
+
+    report("Done")
+    return result
--- a/backend/app/pipeline/turns.py
+++ b/backend/app/pipeline/turns.py
@@ -0,0 +1,60 @@
+import json
+from pathlib import Path
+
+
+def fmt_time(seconds: float) -> str:
+    h, rem = divmod(int(seconds), 3600)
+    m, s = divmod(rem, 60)
+    return f"{h:02d}:{m:02d}:{s:02d}"
+
+
+def load_turns(transcript_json: Path, speaker_map: dict[str, str] | None = None) -> list[dict]:
+    """Parse whisperx json into merged (start, end, speaker, text) turns.
+    speaker_map maps raw labels (SPEAKER_00...) to display names; unmapped labels pass through raw."""
+    data = json.loads(transcript_json.read_text())
+    speaker_map = speaker_map or {}
+    turns = []
+    for seg in data.get("segments", []):
+        raw = seg.get("speaker", "UNKNOWN")
+        speaker = speaker_map.get(raw, raw)
+        text = (seg.get("text") or "").strip()
+        if not text:
+            continue
+        if turns and turns[-1]["speaker"] == speaker:
+            turns[-1]["text"] += " " + text
+            turns[-1]["end"] = seg["end"]
+        else:
+            turns.append({"start": seg["start"], "end": seg["end"], "speaker": speaker, "text": text, "raw_speaker": raw})
+    return turns
+
+
+def distinct_speakers(transcript_json: Path) -> list[dict]:
+    """Return each raw speaker label with a couple of sample lines, for the naming UI."""
+    turns = load_turns(transcript_json)  # no map -> raw labels
+    seen: dict[str, list[dict]] = {}
+    for t in turns:
+        seen.setdefault(t["raw_speaker"], []).append(t)
+    return [
+        {
+            "raw_label": label,
+            "samples": [{"start": t["start"], "text": t["text"][:140]} for t in items[:3]],
+        }
+        for label, items in sorted(seen.items())
+    ]
+
+
+def chunk_turns(turns: list[dict], word_target: int) -> list[list[dict]]:
+    chunks, current, word_count = [], [], 0
+    for turn in turns:
+        current.append(turn)
+        word_count += len(turn["text"].split())
+        if word_count >= word_target:
+            chunks.append(current)
+            current, word_count = [], 0
+    if current:
+        chunks.append(current)
+    return chunks
+
+
+def turns_to_text(turns: list[dict]) -> str:
+    return "\n".join(f"[{fmt_time(t['start'])}] {t['speaker']}: {t['text']}" for t in turns)
--- a/backend/app/routers/init.py
+++ b/backend/app/routers/init.py
--- a/backend/app/routers/diagnostics.py
+++ b/backend/app/routers/diagnostics.py
@@ -0,0 +1,11 @@
+from fastapi import APIRouter
+
+from ..diagnostics import run_diagnostics
+
+router = APIRouter(prefix="/api/diagnostics", tags=["diagnostics"])
+
+
+@router.get("")
+def get_diagnostics():
+    checks = run_diagnostics()
+    return {"ok": all(c["ok"] for c in checks), "checks": checks}
--- a/backend/app/routers/jobs.py
+++ b/backend/app/routers/jobs.py
@@ -0,0 +1,82 @@
+import os
+
+from fastapi import APIRouter, HTTPException, Query
+
+from .. import database as db, config
+
+router = APIRouter(prefix="/api/jobs", tags=["jobs"])
+
+# Canonical job status vocabulary (matches the `jobs` table): queued|running|done|error.
+# The frontend must use these exact strings, not invented ones like "pending"/"completed"/"failed".
+
+
+@router.get("/{job_id}")
+def get_job(job_id: str):
+    job = db.get_job(job_id)
+    if not job:
+        raise HTTPException(404, "Job not found")
+    return job
+
+
+@router.post("/{job_id}/cancel")
+async def cancel_job(job_id: str):
+    job = db.get_job(job_id)
+    if not job:
+        raise HTTPException(404, "Job not found")
+    if job["status"] in ("queued", "running"):
+        # The thread pool executor has no hard-kill hook, so this marks the job
+        # as cancelled; the pipeline itself does not currently check for
+        # cancellation mid-run, so a running job may still finish in the background.
+        db.update_job(job_id, status="error", error="Cancelled by user", error_stage="cancelled")
+    return {"status": "success", "message": f"Job {job_id} cancellation signal sent."}
+
+
+@router.delete("/{job_id}")
+async def delete_job(
+    job_id: str,
+    strategy: str = Query(..., description="One of: 'all', 'artifacts_only', 'none'"),
+):
+    job = db.get_job(job_id)
+    if not job:
+        raise HTTPException(404, "Job not found")
+
+    session_id = job["session_id"]
+    session_row = db.get_conn().execute("SELECT * FROM sessions WHERE id = ?", (session_id,)).fetchone()
+    session = dict(session_row) if session_row else None
+
+    purged = []
+
+    if strategy in ("all", "artifacts_only") and session:
+        for path_key in ("transcript_path",):
+            path = session.get(path_key)
+            if path and os.path.exists(path):
+                os.remove(path)
+                purged.append(path_key)
+        with db.tx() as conn:
+            conn.execute("DELETE FROM notes WHERE session_id = ?", (session_id,))
+            conn.execute("DELETE FROM speakers WHERE session_id = ?", (session_id,))
+        purged.append("notes_and_speakers")
+
+    if strategy == "all" and session:
+        for path_key in ("video_path", "audio_path"):
+            path = session.get(path_key)
+            if path and os.path.exists(path):
+                os.remove(path)
+                purged.append(path_key)
+        with db.tx() as conn:
+            conn.execute("DELETE FROM sessions WHERE id = ?", (session_id,))
+        purged.append("session_record")
+    elif strategy in ("all", "artifacts_only", "none"):
+        with db.tx() as conn:
+            conn.execute("DELETE FROM jobs WHERE id = ?", (job_id,))
+        purged.append("job_record")
+
+    # If a failed job left the session stuck in 'transcribing', roll it back
+    if job["status"] == "error" and session and session["status"] == "transcribing":
+        with db.tx() as conn:
+            conn.execute("UPDATE sessions SET status = 'uploaded' WHERE id = ?", (session_id,))
+
+    if strategy not in ("all", "artifacts_only", "none"):
+        raise HTTPException(400, f"Unknown strategy '{strategy}'")
+
+    return {"status": "success", "message": f"Job {job_id} cleared using strategy: {strategy}", "purged": purged}
--- a/backend/app/routers/models.py
+++ b/backend/app/routers/models.py
@@ -0,0 +1,20 @@
+import requests
+from fastapi import APIRouter
+
+from .. import database as db
+
+router = APIRouter(prefix="/api/models", tags=["models"])
+
+
+@router.get("/ollama")
+def list_ollama_models():
+    """List models already pulled on the configured Ollama host, so the setup
+    wizard can offer a dropdown instead of asking the user to type a tag blind."""
+    settings = db.get_settings()
+    host = settings.get("ollama_host", "http://host.docker.internal:11434")
+    try:
+        resp = requests.get(f"{host.rstrip('/')}/api/tags", timeout=5)
+        resp.raise_for_status()
+        return {"ok": True, "models": [m["name"] for m in resp.json().get("models", [])]}
+    except Exception as e:
+        return {"ok": False, "error": str(e), "models": []}
--- a/backend/app/routers/notes.py
+++ b/backend/app/routers/notes.py
@@ -0,0 +1,48 @@
+from pathlib import Path
+
+from fastapi import APIRouter, HTTPException
+
+from .. import database as db, jobs
+from ..pipeline.turns import load_turns
+from ..pipeline.summarize import summarize_session
+
+router = APIRouter(prefix="/api/sessions/{session_id}/notes", tags=["notes"])
+
+
+@router.post("/generate")
+def generate_notes(session_id: str):
+    session = db.get_conn().execute("SELECT * FROM sessions WHERE id = ?", (session_id,)).fetchone()
+    if not session:
+        raise HTTPException(404, "Session not found")
+    if not session["transcript_path"]:
+        raise HTTPException(400, "Session hasn't been transcribed yet")
+
+    settings = db.get_settings()
+    speaker_map = {r["raw_label"]: r["display_name"] for r in
+                   db.get_conn().execute("SELECT raw_label, display_name FROM speakers WHERE session_id = ?", (session_id,))
+                   if r["display_name"]}
+
+    job_id = db.create_job(session_id, "summarize")
+
+    def run(progress_cb):
+        turns = load_turns(Path(session["transcript_path"]), speaker_map)
+        dm_notes, player_recap = summarize_session(turns, settings, progress_cb=progress_cb)
+        with db.tx() as conn:
+            conn.execute(
+                "INSERT INTO notes (session_id, dm_notes, player_recap, generated_at) VALUES (?, ?, ?, ?) "
+                "ON CONFLICT(session_id) DO UPDATE SET dm_notes=excluded.dm_notes, "
+                "player_recap=excluded.player_recap, generated_at=excluded.generated_at",
+                (session_id, dm_notes, player_recap, db.now()),
+            )
+            conn.execute("UPDATE sessions SET status = 'complete' WHERE id = ?", (session_id,))
+
+    jobs.submit(job_id, run)
+    return {"job_id": job_id}
+
+
+@router.get("")
+def get_notes(session_id: str):
+    row = db.get_conn().execute("SELECT * FROM notes WHERE session_id = ?", (session_id,)).fetchone()
+    if not row:
+        raise HTTPException(404, "Notes not generated yet")
+    return dict(row)
--- a/backend/app/routers/sessions.py
+++ b/backend/app/routers/sessions.py
@@ -0,0 +1,190 @@
+import json
+import os
+import shutil
+from pathlib import Path
+
+from fastapi import APIRouter, UploadFile, File, Form, HTTPException
+
+from .. import database as db, config, jobs
+from ..errors import PipelineError
+from ..pipeline.audio import extract_audio
+from ..pipeline.transcribe import transcribe_and_diarize
+
+
+def _dedup_path(path: Path) -> Path:
+    """Append a counter suffix like ` (1)`, ` (2)` if the file exists."""
+    if not path.exists():
+        return path
+    stem = path.stem
+    suffix = path.suffix
+    parent = path.parent
+    counter = 1
+    while True:
+        new = parent / f"{stem} ({counter}){suffix}"
+        if not new.exists():
+            return new
+        counter += 1
+
+router = APIRouter(prefix="/api/sessions", tags=["sessions"])
+
+# All uploaded/derived files live under config.UPLOAD_DIR / config.AUDIO_DIR /
+# config.TRANSCRIPT_DIR (all under config.DATA_DIR, the /data volume) - never
+# a hardcoded /app/storage path, which isn't backed by any volume the rest of
+# the app agrees on.
+
+
+def _row_to_dict(row) -> dict:
+    return dict(row)
+
+
+@router.get("")
+def list_sessions():
+    rows = db.get_conn().execute("SELECT * FROM sessions ORDER BY created_at DESC").fetchall()
+    sessions = []
+    for r in rows:
+        s = _row_to_dict(r)
+        job = db.get_conn().execute(
+            "SELECT * FROM jobs WHERE session_id = ? ORDER BY created_at DESC LIMIT 1", (s["id"],)
+        ).fetchone()
+        s["latest_job"] = _row_to_dict(job) if job else None
+        sessions.append(s)
+    return sessions
+
+
+@router.get("/{session_id}")
+def get_session(session_id: str):
+    row = db.get_conn().execute("SELECT * FROM sessions WHERE id = ?", (session_id,)).fetchone()
+    if not row:
+        raise HTTPException(404, "Session not found")
+    session = _row_to_dict(row)
+    job = db.get_conn().execute(
+        "SELECT * FROM jobs WHERE session_id = ? ORDER BY created_at DESC LIMIT 1", (session_id,)
+    ).fetchone()
+    session["latest_job"] = _row_to_dict(job) if job else None
+
+    # Stats computed from transcript
+    if session.get("transcript_path"):
+        tp = Path(session["transcript_path"])
+        if tp.exists():
+            try:
+                data = json.loads(tp.read_text())
+                segments = data.get("segments", [])
+                if segments:
+                    session["audio_duration"] = segments[-1]["end"]
+                    session["word_count"] = sum(len(s.get("text", "").split()) for s in segments)
+                    session["language"] = data.get("language")
+            except (json.JSONDecodeError, OSError):
+                pass
+
+    return session
+
+
+def _run_transcription(session_id: str, video_path: Path, audio_path: Path, transcript_path: Path):
+    settings = db.get_settings()
+
+    def run(progress_cb):
+        db.get_conn()  # ensure thread-local connection exists in this worker thread
+        with db.tx() as conn:
+            conn.execute("UPDATE sessions SET status = 'transcribing' WHERE id = ?", (session_id,))
+        progress_cb("Extracting audio...")
+        try:
+            extract_audio(video_path, audio_path)
+            transcribe_and_diarize(
+                audio_path,
+                transcript_path,
+                model_size=settings.get("whisper_model", "medium"),
+                compute_type=settings.get("whisper_compute_type", "int8"),
+                hf_token=settings.get("hf_token", ""),
+                progress_cb=progress_cb,
+            )
+        except PipelineError:
+            with db.tx() as conn:
+                conn.execute("UPDATE sessions SET status = 'uploaded' WHERE id = ?", (session_id,))
+            raise
+        with db.tx() as conn:
+            conn.execute(
+                "UPDATE sessions SET audio_path = ?, transcript_path = ?, status = 'transcribed' WHERE id = ?",
+                (str(audio_path), str(transcript_path), session_id),
+            )
+
+    job_id = db.create_job(session_id, "transcribe")
+    jobs.submit(job_id, run)
+    return job_id
+
+
+@router.post("")
+async def create_session(name: str = Form(...), file: UploadFile = File(None), upload_path: str = Form(None)):
+    session_id = db.new_id()
+
+    if upload_path:
+        src = Path(upload_path)
+        if not src.exists():
+            raise HTTPException(400, f"Uploaded file not found at {upload_path}")
+        safe_filename = os.path.basename(src)
+        video_path = config.UPLOAD_DIR / f"{session_id}_{safe_filename}"
+        shutil.move(str(src), str(video_path))
+    elif file:
+        safe_filename = os.path.basename(file.filename or f"{session_id}")
+        video_path = config.UPLOAD_DIR / f"{session_id}_{safe_filename}"
+        with open(video_path, "wb") as buffer:
+            shutil.copyfileobj(file.file, buffer)
+    else:
+        raise HTTPException(400, "Either a file upload or upload_path is required")
+
+    with db.tx() as conn:
+        conn.execute(
+            "INSERT INTO sessions (id, name, original_filename, video_path, status, created_at) "
+            "VALUES (?, ?, ?, ?, 'uploaded', ?)",
+            (session_id, name, safe_filename, str(video_path), db.now()),
+        )
+
+    audio_path = config.AUDIO_DIR / f"{session_id}.wav"
+    transcript_path = config.TRANSCRIPT_DIR / f"{session_id}.json"
+    job_id = _run_transcription(session_id, video_path, audio_path, transcript_path)
+
+    return {"session_id": session_id, "job_id": job_id}
+
+
+@router.post("/{session_id}/transcribe")
+def retry_transcription(session_id: str):
+    """Re-run transcription for a session, e.g. after a failed job."""
+    row = db.get_conn().execute("SELECT * FROM sessions WHERE id = ?", (session_id,)).fetchone()
+    if not row:
+        raise HTTPException(404, "Session not found")
+    session = _row_to_dict(row)
+    video_path = Path(session["video_path"])
+    audio_path = config.AUDIO_DIR / f"{session_id}.wav"
+    transcript_path = config.TRANSCRIPT_DIR / f"{session_id}.json"
+    job_id = _run_transcription(session_id, video_path, audio_path, transcript_path)
+    return {"job_id": job_id}
+
+
+@router.post("/upload-chunk")
+async def upload_chunk(
+    file: UploadFile = File(...),
+    chunk_index: int = Form(...),
+    total_chunks: int = Form(...),
+    filename: str = Form(...),
+    upload_id: str = Form(...),
+):
+    """Assemble a large upload sent in chunks (bypasses nginx's client_max_body_size
+    for files bigger than that). Returns a filepath under UPLOAD_DIR that can later
+    be referenced when creating a session, for very large recordings."""
+    safe_filename = os.path.basename(filename)
+    temp_dir = config.UPLOAD_DIR / f"tmp_{upload_id}"
+    temp_dir.mkdir(parents=True, exist_ok=True)
+
+    chunk_path = temp_dir / f"chunk_{chunk_index}"
+    with open(chunk_path, "wb") as buffer:
+        shutil.copyfileobj(file.file, buffer)
+
+    if len(os.listdir(temp_dir)) == total_chunks:
+        final_file_path = _dedup_path(config.UPLOAD_DIR / safe_filename)
+        with open(final_file_path, "wb") as final_file:
+            for i in range(total_chunks):
+                with open(temp_dir / f"chunk_{i}", "rb") as chunk_f:
+                    final_file.write(chunk_f.read())
+        shutil.rmtree(temp_dir)
+        return {"status": "completed", "filepath": str(final_file_path), "filename": safe_filename}
+
+    return {"status": "chunk_received", "chunk_index": chunk_index}
--- a/backend/app/routers/settings.py
+++ b/backend/app/routers/settings.py
@@ -0,0 +1,62 @@
+import os
+
+from fastapi import APIRouter
+
+from .. import database as db, config
+
+router = APIRouter(prefix="/api/settings", tags=["settings"])
+
+# Settings are stored as flat key/value rows (see database.py / config.DEFAULT_SETTINGS).
+# This router is a thin pass-through over that table - it must not invent its own
+# schema, or the Settings/Setup pages will silently stop affecting the actual
+# transcription/summarization pipeline, which reads these same flat keys.
+
+# Env var overrides — set these in docker-compose.yml to prefill the setup wizard
+_ENV_OVERRIDES = {
+    "hf_token": "NAT20_HF_TOKEN",
+    "whisper_model": "NAT20_WHISPER_MODEL",
+    "whisper_compute_type": "NAT20_WHISPER_COMPUTE_TYPE",
+    "ollama_host": "NAT20_OLLAMA_HOST",
+    "ollama_model": "NAT20_OLLAMA_MODEL",
+    "api_base_url": "NAT20_API_BASE_URL",
+    "api_key": "NAT20_API_KEY",
+    "api_model": "NAT20_API_MODEL",
+    "chunk_word_target": "NAT20_CHUNK_WORD_TARGET",
+    "world_context": "NAT20_WORLD_CONTEXT",
+    "world_context_path": "NAT20_WORLD_CONTEXT_PATH",
+}
+
+_BOOL_KEYS = {"onboarding_completed"}
+
+
+def _coerce(key: str, value: str):
+    if key in _BOOL_KEYS:
+        return str(value).lower() in ("1", "true", "yes")
+    return value
+
+
+def _merge_with_env(raw: dict) -> dict:
+    """DB values win, env vars fill gaps, DEFAULT_SETTINGS fill remaining gaps."""
+    merged = {**config.DEFAULT_SETTINGS, **raw}
+    for key, env_name in _ENV_OVERRIDES.items():
+        if not raw.get(key, "").strip():
+            val = os.environ.get(env_name)
+            if val is not None:
+                merged[key] = val
+    return merged
+
+
+@router.get("")
+async def get_settings():
+    raw = db.get_settings()
+    merged = _merge_with_env(raw)
+    return {k: _coerce(k, v) for k, v in merged.items()}
+
+
+@router.post("")
+async def update_settings(patch: dict):
+    clean = {k: str(v) for k, v in patch.items() if k in config.DEFAULT_SETTINGS}
+    db.update_settings(clean)
+    raw = db.get_settings()
+    merged = _merge_with_env(raw)
+    return {k: _coerce(k, v) for k, v in merged.items()}
--- a/backend/app/routers/speakers.py
+++ b/backend/app/routers/speakers.py
@@ -0,0 +1,91 @@
+import json
+from pathlib import Path
+
+from fastapi import APIRouter, HTTPException
+from pydantic import BaseModel
+
+from .. import database as db, config
+from ..pipeline.turns import distinct_speakers, load_turns
+
+router = APIRouter(prefix="/api/sessions/{session_id}/speakers", tags=["speakers"])
+
+
+def _get_session_or_404(session_id: str):
+    row = db.get_conn().execute("SELECT * FROM sessions WHERE id = ?", (session_id,)).fetchone()
+    if not row:
+        raise HTTPException(404, "Session not found")
+    return dict(row)
+
+
+@router.get("")
+def get_speakers(session_id: str):
+    session = _get_session_or_404(session_id)
+    if not session["transcript_path"]:
+        raise HTTPException(400, "Session hasn't been transcribed yet")
+
+    speakers = distinct_speakers(Path(session["transcript_path"]))
+
+    existing = {r["raw_label"]: r["display_name"] for r in
+                db.get_conn().execute("SELECT raw_label, display_name FROM speakers WHERE session_id = ?", (session_id,))}
+
+    for sp in speakers:
+        sp["display_name"] = existing.get(sp["raw_label"], "")
+    return speakers
+
+
+class SpeakerName(BaseModel):
+    raw_label: str
+    display_name: str
+
+
+@router.put("")
+def set_speaker_name(session_id: str, body: SpeakerName):
+    _get_session_or_404(session_id)
+    with db.tx() as conn:
+        conn.execute(
+            "INSERT INTO speakers (id, session_id, raw_label, display_name) VALUES (?, ?, ?, ?) "
+            "ON CONFLICT(session_id, raw_label) DO UPDATE SET display_name = excluded.display_name",
+            (db.new_id(), session_id, body.raw_label, body.display_name),
+        )
+    return {"ok": True}
+
+
+class MergeSpeakersBody(BaseModel):
+    target_label: str
+    source_labels: list[str]
+
+
+@router.post("/merge")
+def merge_speakers(session_id: str, body: MergeSpeakersBody):
+    session = _get_session_or_404(session_id)
+    if not session["transcript_path"]:
+        raise HTTPException(400, "Session hasn't been transcribed yet")
+    if body.target_label in body.source_labels:
+        raise HTTPException(400, "Target label cannot also be a source label")
+
+    tp = Path(session["transcript_path"])
+    data = json.loads(tp.read_text())
+    changed = 0
+    for seg in data.get("segments", []):
+        if seg.get("speaker") in body.source_labels:
+            seg["speaker"] = body.target_label
+            changed += 1
+    tp.write_text(json.dumps(data, ensure_ascii=False), encoding="utf-8")
+
+    with db.tx() as conn:
+        for label in body.source_labels:
+            conn.execute("DELETE FROM speakers WHERE session_id = ? AND raw_label = ?", (session_id, label))
+
+    return {"ok": True, "segments_reassigned": changed}
+
+
+@router.get("/turns")
+def get_turns(session_id: str):
+    """Full transcript with display names applied - powers the review/playback screen."""
+    session = _get_session_or_404(session_id)
+    if not session["transcript_path"]:
+        raise HTTPException(400, "Session hasn't been transcribed yet")
+    speaker_map = {r["raw_label"]: r["display_name"] for r in
+                   db.get_conn().execute("SELECT raw_label, display_name FROM speakers WHERE session_id = ?", (session_id,))
+                   if r["display_name"]}
+    return load_turns(Path(session["transcript_path"]), speaker_map)
--- a/backend/requirements.txt
+++ b/backend/requirements.txt
@@ -0,0 +1,24 @@
+fastapi==0.115.0
+uvicorn[standard]==0.30.6
+python-multipart==0.0.9
+requests==2.32.3
+
+# Pin torch/torchaudio explicitly to the cu121 wheels (see --extra-index-url in
+# the Dockerfile). If left unpinned, `pip install whisperx` resolves its
+# loose "torch>=2" constraint against whatever the newest release is
+# (currently a cu128 build) - that's both a much bigger download (duplicate/
+# newer CUDA libs) and known to be ABI-incompatible with the cuDNN version
+# ctranslate2 (a whisperx dependency) expects on older pins. Pinning fixes
+# both the image size and a real "libcudnn_ops_infer.so.8: cannot open
+# shared object file" crash some torch/ctranslate2 combos hit.
+torch==2.3.1
+torchaudio==2.3.1
+
+whisperx==3.3.1
+ctranslate2==4.4.0
+faster-whisper==1.1.0
+pyannote.audio==3.3.2
+nltk==3.9.1
+pandas==2.2.3
+transformers==4.44.2
+matplotlib
--- a/docker-compose.example.yml
+++ b/docker-compose.example.yml
@@ -0,0 +1,33 @@
+# Example Docker Compose — uses a named volume so data persists across rebuilds
+# without needing a specific host path. Copy to docker-compose.yml and customise.
+services:
+  backend:
+    build: ./backend
+    restart: unless-stopped
+    volumes:
+      - app_data:/data
+      - hf_cache:/root/.cache/huggingface
+      - torch_cache:/root/.cache/torch
+
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    ports:
+      - "8000:8000"
+
+  frontend:
+    build: ./frontend
+    restart: unless-stopped
+    depends_on:
+      - backend
+    ports:
+      - "8020:8020"
+
+volumes:
+  app_data:
+  hf_cache:
+  torch_cache:
--- a/frontend/Dockerfile
+++ b/frontend/Dockerfile
@@ -0,0 +1,11 @@
+FROM node:20-alpine AS build
+WORKDIR /app
+COPY package.json ./
+RUN npm install
+COPY . .
+RUN npm run build
+
+FROM nginx:alpine
+COPY --from=build /app/dist /usr/share/nginx/html
+COPY nginx.conf /etc/nginx/conf.d/default.conf
+EXPOSE 8020
--- a/frontend/index.html
+++ b/frontend/index.html
@@ -0,0 +1,14 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="UTF-8" />
+    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+    <title>Nat20 Notes</title>
+    <link rel="preconnect" href="https://fonts.googleapis.com" />
+    <link href="https://fonts.googleapis.com/css2?family=Fraunces:opsz,wght@9..144,400;9..144,600;9..144,700&family=Inter:wght@400;500;600&family=IBM+Plex+Mono:wght@400;500&display=swap" rel="stylesheet" />
+  </head>
+  <body class="bg-deep text-ink font-body">
+    <div id="root"></div>
+    <script type="module" src="/src/main.tsx"></script>
+  </body>
+</html>
--- a/frontend/nginx.conf
+++ b/frontend/nginx.conf
@@ -0,0 +1,17 @@
+server {
+    listen 8020;
+
+    client_max_body_size 500M;
+
+    location /api/ {
+        proxy_pass http://backend:8000/api/;
+        proxy_set_header Host $host;
+        proxy_read_timeout 3600s;
+        proxy_send_timeout 3600s;
+    }
+
+    location / {
+        root /usr/share/nginx/html;
+        try_files $uri /index.html;
+    }
+}
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -0,0 +1,26 @@
+{
+  "name": "session-notes-frontend",
+  "private": true,
+  "version": "0.1.0",
+  "type": "module",
+  "scripts": {
+    "dev": "vite",
+    "build": "vite build",
+    "preview": "vite preview"
+  },
+  "dependencies": {
+    "react": "^18.3.1",
+    "react-dom": "^18.3.1",
+    "react-router-dom": "^6.26.2"
+  },
+  "devDependencies": {
+    "@types/react": "^18.3.5",
+    "@types/react-dom": "^18.3.0",
+    "@vitejs/plugin-react": "^4.3.1",
+    "autoprefixer": "^10.4.20",
+    "postcss": "^8.4.47",
+    "tailwindcss": "^3.4.11",
+    "typescript": "^5.6.2",
+    "vite": "^5.4.6"
+  }
+}
--- a/frontend/postcss.config.js
+++ b/frontend/postcss.config.js
@@ -0,0 +1,6 @@
+export default {
+  plugins: {
+    tailwindcss: {},
+    autoprefixer: {},
+  },
+};
--- a/frontend/src/App.tsx
+++ b/frontend/src/App.tsx
@@ -0,0 +1,59 @@
+import { useState, useEffect } from 'react';
+import { Routes, Route, Navigate } from 'react-router-dom';
+import Setup from './pages/Setup';
+import Sessions from './pages/Sessions';
+import SessionDetail from './pages/SessionDetail';
+import Speakers from './pages/Speakers';
+import Notes from './pages/Notes';
+import Settings from './pages/Settings';
+import Diagnostics from './pages/Diagnostics';
+import { api } from './api';
+
+function App() {
+  const [loading, setLoading] = useState(true);
+  const [isConfigured, setIsConfigured] = useState(false);
+
+  const checkSettings = () => {
+    api.getSettings()
+      .then(data => {
+        setIsConfigured(!!(data && data.onboarding_completed));
+      })
+      .catch(() => setIsConfigured(false))
+      .finally(() => setLoading(false));
+  };
+
+  useEffect(() => {
+    checkSettings();
+  }, []);
+
+  if (loading) {
+    return <div className="min-h-screen bg-slate-900 text-white flex items-center justify-center">Loading...</div>;
+  }
+
+  return (
+    <Routes>
+      {/* If not configured, force them to the Setup wizard */}
+      <Route
+        path="/setup"
+        element={!isConfigured ? <Setup onComplete={checkSettings} /> : <Navigate to="/" />}
+      />
+
+      {/* If configured, serve the main sessions workspace */}
+      {isConfigured ? (
+        <>
+          <Route path="/" element={<Sessions />} />
+          <Route path="/sessions/:id" element={<SessionDetail />} />
+          <Route path="/sessions/:id/speakers" element={<Speakers />} />
+          <Route path="/sessions/:id/notes" element={<Notes />} />
+          <Route path="/settings" element={<Settings />} />
+          <Route path="/diagnostics" element={<Diagnostics />} />
+          <Route path="*" element={<Navigate to="/" />} />
+        </>
+      ) : (
+        <Route path="*" element={<Navigate to="/setup" />} />
+      )}
+    </Routes>
+  );
+}
+
+export default App;
--- a/frontend/src/api.ts
+++ b/frontend/src/api.ts
@@ -0,0 +1,190 @@
+const BASE_URL = '/api';
+
+// Flat settings schema - must mirror backend/app/config.py's DEFAULT_SETTINGS
+// exactly. This is the single source of truth; don't reintroduce a nested
+// shape here without updating config.py, routers/settings.py, Setup.tsx and
+// Settings.tsx together.
+export interface AppSettings {
+  onboarding_completed: boolean;
+  whisper_model: string;
+  whisper_compute_type: string;
+  hf_token: string;
+  llm_mode: 'ollama' | 'api';
+  ollama_host: string;
+  ollama_model: string;
+  api_base_url: string;
+  api_key: string;
+  api_model: string;
+  chunk_word_target: string;
+  world_context: string;
+  world_context_path: string;
+}
+
+export type JobStatus = 'queued' | 'running' | 'done' | 'error';
+
+export interface Job {
+  id: string;
+  session_id: string;
+  job_type: 'transcribe' | 'summarize';
+  status: JobStatus;
+  progress: string | null;
+  error: string | null;
+  error_stage: string | null;
+  created_at: number;
+  updated_at: number;
+}
+
+export type DeleteStrategy = 'all' | 'artifacts_only' | 'none';
+
+export const jobApi = {
+  getJob: async (jobId: string): Promise<Job> => {
+    const res = await fetch(`${BASE_URL}/jobs/${jobId}`);
+    if (!res.ok) throw new Error('Failed to fetch job');
+    return res.json();
+  },
+  cancelJob: async (jobId: string): Promise<any> => {
+    const res = await fetch(`${BASE_URL}/jobs/${jobId}/cancel`, { method: 'POST' });
+    return res.json();
+  },
+  deleteJob: async (jobId: string, strategy: DeleteStrategy): Promise<any> => {
+    const res = await fetch(`${BASE_URL}/jobs/${jobId}?strategy=${strategy}`, { method: 'DELETE' });
+    if (!res.ok) throw new Error('Failed to delete job');
+    return res.json();
+  },
+};
+
+export const api = {
+  // Settings
+  getSettings: async (): Promise<AppSettings> => {
+    const res = await fetch(`${BASE_URL}/settings`);
+    return res.json();
+  },
+  updateSettings: async (settings: Partial<AppSettings>): Promise<AppSettings> => {
+    const res = await fetch(`${BASE_URL}/settings`, {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify(settings),
+    });
+    return res.json();
+  },
+  // kept as an alias since some pages call it by this name
+  saveSettings: async (settings: Partial<AppSettings>): Promise<AppSettings> => {
+    return api.updateSettings(settings);
+  },
+
+  // Sessions
+  listSessions: async (): Promise<any[]> => {
+    const res = await fetch(`${BASE_URL}/sessions`);
+    return res.json();
+  },
+  createSession: async (name: string, file: File, onProgress?: (pct: number) => void): Promise<{ session_id: string; job_id: string }> => {
+    const CHUNK_THRESHOLD = 90 * 1024 * 1024;
+    if (file.size > CHUNK_THRESHOLD) {
+      const result = await api.uploadFileInChunks(file, onProgress || (() => {}));
+      const formData = new FormData();
+      formData.append('name', name);
+      formData.append('upload_path', result.filepath);
+      const res = await fetch(`${BASE_URL}/sessions`, { method: 'POST', body: formData });
+      if (!res.ok) throw new Error('Failed to create session');
+      return res.json();
+    }
+    const formData = new FormData();
+    formData.append('name', name);
+    formData.append('file', file);
+    const res = await fetch(`${BASE_URL}/sessions`, { method: 'POST', body: formData });
+    if (!res.ok) throw new Error('Failed to create session');
+    return res.json();
+  },
+  getSession: async (sessionId: string): Promise<any> => {
+    const res = await fetch(`${BASE_URL}/sessions/${sessionId}`);
+    if (!res.ok) throw new Error('Failed to fetch session');
+    return res.json();
+  },
+  retryTranscription: async (sessionId: string): Promise<{ job_id: string }> => {
+    const res = await fetch(`${BASE_URL}/sessions/${sessionId}/transcribe`, { method: 'POST' });
+    if (!res.ok) throw new Error('Failed to start transcription');
+    return res.json();
+  },
+  audioUrl: (sessionId: string): string => `${BASE_URL}/sessions/${sessionId}/audio`,
+
+  // Speakers
+  getSpeakers: async (sessionId: string): Promise<any[]> => {
+    const res = await fetch(`${BASE_URL}/sessions/${sessionId}/speakers`);
+    return res.json();
+  },
+  getTurns: async (sessionId: string): Promise<any[]> => {
+    const res = await fetch(`${BASE_URL}/sessions/${sessionId}/speakers/turns`);
+    return res.json();
+  },
+  setSpeakerName: async (sessionId: string, raw_label: string, display_name: string): Promise<void> => {
+    const res = await fetch(`${BASE_URL}/sessions/${sessionId}/speakers`, {
+      method: 'PUT',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ raw_label, display_name }),
+    });
+    if (!res.ok) throw new Error('Failed to save speaker name');
+  },
+
+  // Notes
+  getNotes: async (sessionId: string): Promise<any> => {
+    const res = await fetch(`${BASE_URL}/sessions/${sessionId}/notes`);
+    if (!res.ok) throw new Error('Notes not generated yet');
+    return res.json();
+  },
+  generateNotes: async (sessionId: string): Promise<{ job_id: string }> => {
+    const res = await fetch(`${BASE_URL}/sessions/${sessionId}/notes/generate`, { method: 'POST' });
+    if (!res.ok) throw new Error('Failed to start note generation');
+    return res.json();
+  },
+
+  // Diagnostics
+  getDiagnostics: async (): Promise<{ ok: boolean; checks: any[] }> => {
+    const res = await fetch(`${BASE_URL}/diagnostics`);
+    if (!res.ok) throw new Error('Failed to reach backend');
+    return res.json();
+  },
+
+  // Jobs (re-exported here so pages that only import `api` still work)
+  deleteJob: jobApi.deleteJob,
+  cancelJob: jobApi.cancelJob,
+  getJob: jobApi.getJob,
+
+  // Large-file (chunked) upload, used automatically by createSession()
+  // when the file exceeds 90MB. Sends up to 3 chunks concurrently.
+  uploadFileInChunks: async (
+    file: File,
+    onProgress: (percent: number) => void
+  ): Promise<any> => {
+    const CHUNK_SIZE = 1024 * 1024 * 15;
+    const CONCURRENCY = 3;
+    const totalChunks = Math.ceil(file.size / CHUNK_SIZE);
+    const uploadId = crypto.randomUUID();
+
+    const uploadOne = async (chunkIndex: number): Promise<any> => {
+      const start = chunkIndex * CHUNK_SIZE;
+      const end = Math.min(start + CHUNK_SIZE, file.size);
+      const chunk = file.slice(start, end);
+      const formData = new FormData();
+      formData.append('file', chunk, file.name);
+      formData.append('chunk_index', chunkIndex.toString());
+      formData.append('total_chunks', totalChunks.toString());
+      formData.append('filename', file.name);
+      formData.append('upload_id', uploadId);
+      const res = await fetch(`${BASE_URL}/sessions/upload-chunk`, { method: 'POST', body: formData });
+      if (!res.ok) throw new Error(`Failed to upload chunk ${chunkIndex}`);
+      return res.json();
+    };
+
+    let lastResult: any = null;
+    for (let i = 0; i < totalChunks; i += CONCURRENCY) {
+      const batch = [];
+      for (let j = i; j < Math.min(i + CONCURRENCY, totalChunks); j++) {
+        batch.push(uploadOne(j));
+      }
+      const results = await Promise.all(batch);
+      lastResult = results[results.length - 1];
+      onProgress(Math.round((Math.min(i + CONCURRENCY, totalChunks) / totalChunks) * 100));
+    }
+    return lastResult;
+  },
+};
--- a/frontend/src/components/SessionReel.tsx
+++ b/frontend/src/components/SessionReel.tsx
@@ -0,0 +1,66 @@
+import { useMemo, useRef } from "react";
+
+const SPEAKER_COLORS = ["#C9A227", "#5F7A61", "#4E6A87", "#7A5670", "#B4523A", "#8A8F5C"];
+
+function colorFor(label: string, palette: Map<string, string>) {
+  if (!palette.has(label)) {
+    palette.set(label, SPEAKER_COLORS[palette.size % SPEAKER_COLORS.length]);
+  }
+  return palette.get(label)!;
+}
+
+export function SessionReel({
+  turns,
+  duration,
+  onSeek,
+  currentTime,
+}: {
+  turns: { start: number; end: number; raw_speaker: string }[];
+  duration: number;
+  onSeek: (t: number) => void;
+  currentTime: number;
+}) {
+  const palette = useMemo(() => new Map<string, string>(), []);
+  const trackRef = useRef<HTMLDivElement>(null);
+
+  const valid = duration > 0 && isFinite(duration);
+
+  const handleClick = (e: React.MouseEvent) => {
+    if (!trackRef.current || !valid) return;
+    const rect = trackRef.current.getBoundingClientRect();
+    const frac = (e.clientX - rect.left) / rect.width;
+    onSeek(Math.max(0, Math.min(duration, frac * duration)));
+  };
+
+  const playheadPct = valid ? (currentTime / duration) * 100 : 0;
+
+  return (
+    <div
+      ref={trackRef}
+      onClick={handleClick}
+      className={"relative h-14 rounded-md border overflow-hidden " + (valid ? "bg-deep border-white/10 cursor-pointer" : "bg-panel2 border-white/5")}
+      role="slider"
+      aria-label="Session timeline"
+    >
+      {valid &&
+        turns.map((t, i) => {
+          const left = (t.start / duration) * 100;
+          const width = Math.max(((t.end - t.start) / duration) * 100, 0.15);
+          return (
+            <div
+              key={i}
+              className="absolute top-1.5 bottom-1.5 rounded-sm opacity-70 hover:opacity-100 transition"
+              style={{ left: `${left}%`, width: `${width}%`, background: colorFor(t.raw_speaker, palette) }}
+              title={`${t.raw_speaker} @ ${Math.floor(t.start)}s`}
+            />
+          );
+        })}
+      <div
+        className="absolute top-0 bottom-0 w-px bg-ink shadow-[0_0_6px_1px_rgba(237,230,214,0.6)]"
+        style={{ left: `${playheadPct}%` }}
+      />
+    </div>
+  );
+}
+
+export { colorFor, SPEAKER_COLORS };
--- a/frontend/src/main.tsx
+++ b/frontend/src/main.tsx
@@ -0,0 +1,13 @@
+import React from "react";
+import ReactDOM from "react-dom/client";
+import { BrowserRouter } from "react-router-dom";
+import App from "./App";
+import "./styles.css";
+
+ReactDOM.createRoot(document.getElementById("root")!).render(
+  <React.StrictMode>
+    <BrowserRouter>
+      <App />
+    </BrowserRouter>
+  </React.StrictMode>
+);
--- a/frontend/src/pages/Diagnostics.tsx
+++ b/frontend/src/pages/Diagnostics.tsx
@@ -0,0 +1,55 @@
+import { useEffect, useState } from "react";
+import { Link } from "react-router-dom";
+import { api } from "../api";
+
+const LABELS: Record<string, string> = {
+  ffmpeg: "ffmpeg",
+  gpu: "GPU / CUDA",
+  huggingface_token: "HuggingFace token",
+  llm_backend: "Summarization backend",
+  disk_space: "Disk space",
+};
+
+export default function Diagnostics() {
+  const [result, setResult] = useState<any>(null);
+  const [loading, setLoading] = useState(false);
+
+  const run = async () => {
+    setLoading(true);
+    const r = await api.getDiagnostics().catch((e) => ({ ok: false, checks: [], _error: e.message }));
+    setResult(r);
+    setLoading(false);
+  };
+
+  useEffect(() => {
+    run();
+  }, []);
+
+  return (
+    <div className="max-w-xl mx-auto py-16 px-4">
+      <Link to="/" className="text-sm text-ink/40 hover:text-brass">&larr; Back</Link>
+      <h1 className="font-display text-3xl mt-4 mb-2">System check</h1>
+      <p className="text-ink/60 mb-6">Run this before your first session, or any time something isn't working.</p>
+
+      <button className="btn-secondary mb-6" onClick={run} disabled={loading}>
+        {loading ? "Checking..." : "Run checks again"}
+      </button>
+
+      {result?._error && (
+        <div className="card p-4 text-ember">Couldn't reach the backend: {result._error}</div>
+      )}
+
+      <div className="space-y-3">
+        {result?.checks?.map((c: any) => (
+          <div key={c.name} className="card p-4 flex gap-3">
+            <span className={`w-2 h-2 rounded-full mt-1.5 shrink-0 ${c.ok ? "bg-moss" : "bg-ember"}`} />
+            <div>
+              <div className="font-medium">{LABELS[c.name] || c.name}</div>
+              <div className="text-sm text-ink/60 whitespace-pre-wrap">{c.message}</div>
+            </div>
+          </div>
+        ))}
+      </div>
+    </div>
+  );
+}
--- a/frontend/src/pages/Notes.tsx
+++ b/frontend/src/pages/Notes.tsx
@@ -0,0 +1,35 @@
+import { useEffect, useState } from "react";
+import { Link, useParams } from "react-router-dom";
+import { api } from "../api";
+
+export default function Notes() {
+  const { id } = useParams();
+  const [notes, setNotes] = useState<any>(null);
+  const [tab, setTab] = useState<"dm" | "player">("dm");
+
+  useEffect(() => {
+    api.getNotes(id!).then(setNotes);
+  }, [id]);
+
+  if (!notes) return null;
+
+  return (
+    <div className="max-w-3xl mx-auto py-16 px-4">
+      <Link to={`/sessions/${id}`} className="text-sm text-ink/40 hover:text-brass">&larr; Back</Link>
+      <h1 className="font-display text-3xl mt-4 mb-6">Session notes</h1>
+
+      <div className="flex gap-2 mb-6">
+        <button className={tab === "dm" ? "btn-primary" : "btn-secondary"} onClick={() => setTab("dm")}>
+          DM notes
+        </button>
+        <button className={tab === "player" ? "btn-primary" : "btn-secondary"} onClick={() => setTab("player")}>
+          Player recap
+        </button>
+      </div>
+
+      <div className="card p-6 whitespace-pre-wrap leading-relaxed">
+        {tab === "dm" ? notes.dm_notes : notes.player_recap}
+      </div>
+    </div>
+  );
+}
--- a/frontend/src/pages/SessionDetail.tsx
+++ b/frontend/src/pages/SessionDetail.tsx
@@ -0,0 +1,257 @@
+import { useEffect, useState, useCallback, useRef } from 'react';
+import { Link, useNavigate, useParams } from 'react-router-dom';
+import { api, Job } from '../api';
+
+function Toast({ message, onDone }: { message: string; onDone: () => void }) {
+  useEffect(() => { const t = setTimeout(onDone, 2500); return () => clearTimeout(t); }, [onDone]);
+  return (
+    <div className="fixed top-4 left-1/2 -translate-x-1/2 z-50 bg-emerald-700 text-white px-5 py-3 rounded-xl shadow-2xl text-sm font-medium">
+      {message}
+    </div>
+  );
+}
+
+function formatDuration(seconds: number): string {
+  const h = Math.floor(seconds / 3600);
+  const m = Math.floor((seconds % 3600) / 60);
+  const s = Math.floor(seconds % 60);
+  if (h > 0) return `${h}:${m.toString().padStart(2, '0')}:${s.toString().padStart(2, '0')}`;
+  return `${m}:${s.toString().padStart(2, '0')}`;
+}
+
+function ElapsedTimer({ createdAt, endedAt }: { createdAt: number; endedAt?: number }) {
+  const [elapsed, setElapsed] = useState(0);
+  useEffect(() => {
+    if (endedAt !== undefined) {
+      setElapsed(Math.floor(endedAt - createdAt));
+      return;
+    }
+    const tick = () => setElapsed(Math.floor(Date.now() / 1000 - createdAt));
+    tick();
+    const id = setInterval(tick, 1000);
+    return () => clearInterval(id);
+  }, [createdAt, endedAt]);
+  return <span className="font-mono text-xs">{formatDuration(elapsed)}</span>;
+}
+
+const STATUS_LABEL: Record<string, string> = {
+  uploaded: 'Ready to transcribe',
+  transcribing: 'Transcribing...',
+  transcribed: 'Ready to name speakers',
+  complete: 'Notes ready',
+};
+
+function JobStatusCard({ job, session, onRefresh, onDeleted }: { job: Job; session?: any; onRefresh: () => void; onDeleted: (strategy: string) => void }) {
+  const [showDeleteModal, setShowDeleteModal] = useState(false);
+  const [busy, setBusy] = useState(false);
+
+  const handleCancel = async () => {
+    if (!confirm('Are you sure you want to halt this job midway?')) return;
+    setBusy(true);
+    await api.cancelJob(job.id);
+    onRefresh();
+    setBusy(false);
+  };
+
+  const executeDeletion = async (strategy: 'all' | 'artifacts_only' | 'none') => {
+    setBusy(true);
+    await api.deleteJob(job.id, strategy);
+    setShowDeleteModal(false);
+    setBusy(false);
+    onDeleted(strategy);
+  };
+
+  const badgeClass =
+    job.status === 'done'
+      ? 'bg-emerald-500/20 text-emerald-400 border border-emerald-500/30'
+      : job.status === 'error'
+      ? 'bg-rose-500/20 text-rose-400 border border-rose-500/30'
+      : 'bg-amber-500/20 text-amber-400 border border-amber-500/30 animate-pulse';
+
+  return (
+    <div className="card p-5">
+      <div className="flex items-center justify-between mb-3">
+        <h3 className="font-semibold text-lg">{job.job_type === 'transcribe' ? 'Transcription' : 'Note generation'}</h3>
+        <span className={`px-2.5 py-1 text-xs font-bold rounded-full uppercase ${badgeClass}`}>
+          {job.status === 'error' ? 'Failed' : job.status}
+        </span>
+      </div>
+
+      {job.status === 'error' ? (
+        <div className="bg-rose-950/40 border border-rose-900/50 text-rose-300 text-sm p-3 rounded mb-3 font-mono">
+          <strong>Error{job.error_stage ? ` [${job.error_stage}]` : ''}:</strong> {job.error || 'Unexpected error.'}
+        </div>
+      ) : (
+        <p className="text-sm text-ink/60 mb-3">{job.progress || 'Waiting...'}</p>
+      )}
+
+      {job.status !== 'queued' && job.created_at && (
+        <div className="flex items-center gap-1 text-ink/40 mb-3">
+          <svg className="w-3.5 h-3.5" fill="none" viewBox="0 0 24 24" stroke="currentColor" strokeWidth={2}><path strokeLinecap="round" strokeLinejoin="round" d="M12 8v4l3 3m6-3a9 9 0 11-18 0 9 9 0 0118 0z"/></svg>
+          <ElapsedTimer
+            createdAt={job.created_at}
+            endedAt={job.status === 'done' || job.status === 'error' ? job.updated_at : undefined}
+          />
+        </div>
+      )}
+
+      {session?.audio_duration && (
+        <div className="text-xs text-ink/40 mb-3 font-mono">
+          Audio {formatDuration(session.audio_duration)}
+          {session.word_count ? ` \u2022 ${session.word_count.toLocaleString()} words` : ''}
+          {session.language ? ` \u2022 ${session.language}` : ''}
+        </div>
+      )}
+
+      <div className="flex gap-3">
+        {(job.status === 'queued' || job.status === 'running') && (
+          <button disabled={busy} onClick={handleCancel} className="btn-secondary text-sm">
+            Cancel
+          </button>
+        )}
+        {job.status === 'error' && job.job_type === 'transcribe' && (
+          <button
+            disabled={busy}
+            onClick={async () => { setBusy(true); await api.retryTranscription(job.session_id); setBusy(false); onRefresh(); }}
+            className="btn-primary text-sm"
+          >
+            Restart transcription
+          </button>
+        )}
+        <button disabled={busy} onClick={() => setShowDeleteModal(true)} className="btn-secondary text-sm ml-auto">
+          Delete...
+        </button>
+      </div>
+
+      {showDeleteModal && (
+        <div className="fixed inset-0 bg-black/60 flex items-center justify-center p-4 z-50">
+          <div className="bg-panel border border-white/10 max-w-md w-full rounded-xl p-6 shadow-2xl">
+            <h4 className="text-lg font-bold mb-2">Clean up session data?</h4>
+            <p className="text-sm text-ink/60 mb-6 leading-relaxed">
+              Choose what to remove along with this job's tracking record.
+            </p>
+            <div className="space-y-2">
+              <button onClick={() => executeDeletion('all')} className="w-full text-left card p-3 text-sm">
+                🗑️ <strong>Wipe everything</strong>
+                <span className="block text-xs text-ink/40 mt-0.5">Session record, notes, transcript, and the source audio/video.</span>
+              </button>
+              <button onClick={() => executeDeletion('artifacts_only')} className="w-full text-left card p-3 text-sm">
+                📝 <strong>Keep source file</strong>
+                <span className="block text-xs text-ink/40 mt-0.5">Clears notes/transcript but keeps the original recording.</span>
+              </button>
+              <button onClick={() => executeDeletion('none')} className="w-full text-left card p-3 text-sm">
+                ❌ <strong>Just clear this job</strong>
+                <span className="block text-xs text-ink/40 mt-0.5">Only removes the job tracking row. No files touched.</span>
+              </button>
+            </div>
+            <div className="mt-5 flex justify-end">
+              <button onClick={() => setShowDeleteModal(false)} className="text-sm text-ink/50 hover:text-ink">
+                Cancel
+              </button>
+            </div>
+          </div>
+        </div>
+      )}
+    </div>
+  );
+}
+
+export default function SessionDetail() {
+  const { id } = useParams();
+  const nav = useNavigate();
+  const [session, setSession] = useState<any>(null);
+  const [job, setJob] = useState<Job | null>(null);
+  const [toast, setToast] = useState<string | null>(null);
+
+  const load = useCallback(() => {
+    if (!id) return;
+    api.getSession(id).then((s) => {
+      setSession(s);
+      setJob(s.latest_job || null);
+    });
+  }, [id]);
+
+  useEffect(load, [load]);
+
+  // Poll while a job is actively running so status/progress stay live.
+  useEffect(() => {
+    if (!job || (job.status !== 'queued' && job.status !== 'running')) return;
+    const interval = setInterval(load, 2500);
+    return () => clearInterval(interval);
+  }, [job, load]);
+
+  // Preload audio as soon as transcription finishes so it's buffered for the speakers page
+  useEffect(() => {
+    if (session?.status !== 'transcribed' || !id) return;
+    const link = document.createElement('link');
+    link.rel = 'preload';
+    link.as = 'audio';
+    link.href = api.audioUrl(id);
+    document.head.appendChild(link);
+    return () => link.remove();
+  }, [session?.status, id]);
+
+  // Auto-redirect to speakers page once transcription completes
+  const autoRedirected = useRef(false);
+  useEffect(() => {
+    if (session?.status === 'transcribed') {
+      if (!autoRedirected.current) {
+        autoRedirected.current = true;
+        nav(`/sessions/${id}/speakers`);
+      }
+    } else {
+      autoRedirected.current = false;
+    }
+  }, [session?.status, id, nav]);
+
+  const retryTranscription = async () => {
+    if (!id) return;
+    await api.retryTranscription(id);
+    load();
+  };
+
+  const generateNotes = async () => {
+    if (!id) return;
+    await api.generateNotes(id);
+    load();
+  };
+
+  if (!session) return null;
+
+  return (
+    <div className="max-w-2xl mx-auto py-16 px-4">
+      {toast && <Toast message={toast} onDone={() => setToast(null)} />}
+      <Link to="/" className="text-sm text-ink/40 hover:text-brass">&larr; Back</Link>
+      <h1 className="font-display text-3xl mt-4 mb-1">{session.name}</h1>
+      <p className="text-ink/40 font-mono text-xs mb-6">{session.original_filename}</p>
+      <p className="text-brass mb-8">{STATUS_LABEL[session.status] ?? session.status}</p>
+
+      {job && (
+        <div className="mb-8">
+          <JobStatusCard job={job} session={session} onRefresh={load} onDeleted={(strategy) => {
+            if (strategy === 'all') {
+              setToast('Session wiped');
+              setTimeout(() => nav('/'), 600);
+            } else if (strategy === 'artifacts_only') {
+              setToast('Transcript and notes cleared — source file kept');
+              load();
+            } else {
+              setToast('Job record removed');
+              load();
+            }
+          }} />
+        </div>
+      )}
+
+      {session.status === 'uploaded' && !job && (
+        <button className="btn-primary" onClick={retryTranscription}>Start transcription</button>
+      )}
+      {session.status === 'complete' && (
+        <div className="mt-4 flex gap-3">
+          <button className="btn-secondary" onClick={generateNotes}>Regenerate notes</button>
+          <Link to={`/sessions/${id}/notes`} className="btn-primary">View notes</Link>
+        </div>
+      )}
+    </div>
+  );
+}
--- a/frontend/src/pages/Sessions.tsx
+++ b/frontend/src/pages/Sessions.tsx
@@ -0,0 +1,95 @@
+import { useEffect, useState, useCallback } from "react";
+import { Link, useNavigate } from "react-router-dom";
+import { api } from "../api";
+
+const STATUS_LABEL: Record<string, string> = {
+  uploaded: "Ready to transcribe",
+  transcribing: "Transcribing...",
+  transcribed: "Ready to name speakers",
+  complete: "Notes ready",
+};
+
+export default function Sessions() {
+  const [sessions, setSessions] = useState<any[]>([]);
+  const [name, setName] = useState("");
+  const [file, setFile] = useState<File | null>(null);
+  const [uploadProgress, setUploadProgress] = useState<number | null>(null);
+  const nav = useNavigate();
+
+  const refresh = useCallback(() => api.listSessions().then(setSessions), []);
+  useEffect(() => {
+    refresh();
+  }, [refresh]);
+
+  // Poll while any session is actively processing
+  useEffect(() => {
+    const hasActive = sessions.some(s => s.status === 'transcribing');
+    if (!hasActive) return;
+    const interval = setInterval(refresh, 2500);
+    return () => clearInterval(interval);
+  }, [sessions, refresh]);
+
+  const upload = async () => {
+    if (!file || !name) return;
+    setUploadProgress(0);
+    const { session_id } = await api.createSession(name, file, setUploadProgress);
+    setUploadProgress(null);
+    nav(`/sessions/${session_id}`);
+  };
+
+  return (
+    <div className="max-w-3xl mx-auto py-16 px-4">
+      <div className="eyebrow mb-2">Nat20 Notes</div>
+      <h1 className="font-display text-4xl mb-8">Your Transcriptions</h1>
+
+      <div className="card p-6 mb-8">
+        <h2 className="font-display text-xl mb-4">New transcription</h2>
+        <div className="flex flex-col gap-3">
+          <input
+            className="input"
+            placeholder="Session name, e.g. Session 1 — The Frozen Crypt"
+            value={name}
+            onChange={(e) => setName(e.target.value)}
+          />
+          <input
+            type="file"
+            accept="audio/*,video/*"
+            onChange={(e) => setFile(e.target.files?.[0] ?? null)}
+            className="text-sm text-ink/70"
+          />
+          <button className="btn-primary self-start" disabled={!file || !name || uploadProgress !== null} onClick={upload}>
+            {uploadProgress !== null ? `Uploading ${uploadProgress}%` : "Upload"}
+          </button>
+        </div>
+      </div>
+
+      <div className="space-y-3">
+        {sessions.length === 0 && (
+          <p className="text-ink/40">No sessions yet — upload a recording to get started.</p>
+        )}
+        {sessions.map((s) => {
+          const job = s.latest_job;
+          const progress = job?.progress;
+          const displayStatus = s.status === 'transcribing' && progress
+            ? progress
+            : (STATUS_LABEL[s.status] ?? s.status);
+          const statusColor = job?.status === 'error' ? 'text-rose-400' : 'text-brass';
+          return (
+            <Link key={s.id} to={`/sessions/${s.id}`} className="card p-4 flex items-center justify-between hover:border-brass/40 border border-transparent block">
+              <div>
+                <div className="font-medium">{s.name}</div>
+                <div className="text-xs text-ink/40 font-mono">{s.original_filename}</div>
+              </div>
+              <div className={"text-sm " + statusColor}>{displayStatus}</div>
+            </Link>
+          );
+        })}
+      </div>
+
+      <div className="mt-10 flex gap-4">
+        <Link to="/diagnostics" className="text-sm text-ink/40 hover:text-brass">System check</Link>
+        <Link to="/settings" className="text-sm text-ink/40 hover:text-brass">Settings</Link>
+      </div>
+    </div>
+  );
+}
--- a/frontend/src/pages/Settings.tsx
+++ b/frontend/src/pages/Settings.tsx
@@ -0,0 +1,68 @@
+import { useEffect, useState } from "react";
+import { Link } from "react-router-dom";
+import { api } from "../api";
+
+export default function Settings() {
+  const [settings, setSettings] = useState<any>(null);
+  const [saved, setSaved] = useState(false);
+
+  useEffect(() => {
+    api.getSettings().then(setSettings);
+  }, []);
+
+  if (!settings) return null;
+
+  const set = (k: string, v: any) => setSettings({ ...settings, [k]: v });
+
+  const save = async () => {
+    await api.updateSettings(settings);
+    setSaved(true);
+    setTimeout(() => setSaved(false), 2000);
+  };
+
+  return (
+    <div className="max-w-xl mx-auto py-16 px-4">
+      <Link to="/" className="text-sm text-ink/40 hover:text-brass">&larr; Back</Link>
+      <h1 className="font-display text-3xl mt-4 mb-8">Settings</h1>
+
+      <div className="card p-6 space-y-4 mb-6">
+        <h2 className="eyebrow">Transcription</h2>
+        <select className="input w-full" value={settings.whisper_model} onChange={(e) => set("whisper_model", e.target.value)}>
+          <option value="tiny">Tiny</option>
+          <option value="small">Small</option>
+          <option value="medium">Medium</option>
+          <option value="large-v3">Large v3</option>
+        </select>
+        <input className="input w-full" type="password" placeholder="HuggingFace token" value={settings.hf_token} onChange={(e) => set("hf_token", e.target.value)} />
+      </div>
+
+      <div className="card p-6 space-y-4 mb-6">
+        <h2 className="eyebrow">Summarization</h2>
+        <div className="flex gap-3">
+          <button className={settings.llm_mode === "ollama" ? "btn-primary" : "btn-secondary"} onClick={() => set("llm_mode", "ollama")}>Local (Ollama)</button>
+          <button className={settings.llm_mode === "api" ? "btn-primary" : "btn-secondary"} onClick={() => set("llm_mode", "api")}>Hosted API</button>
+        </div>
+        {settings.llm_mode === "ollama" ? (
+          <>
+            <input className="input w-full" value={settings.ollama_host} onChange={(e) => set("ollama_host", e.target.value)} />
+            <input className="input w-full" value={settings.ollama_model} onChange={(e) => set("ollama_model", e.target.value)} />
+          </>
+        ) : (
+          <>
+            <input className="input w-full" value={settings.api_base_url} onChange={(e) => set("api_base_url", e.target.value)} />
+            <input className="input w-full" type="password" value={settings.api_key} onChange={(e) => set("api_key", e.target.value)} />
+            <input className="input w-full" value={settings.api_model} onChange={(e) => set("api_model", e.target.value)} />
+          </>
+        )}
+      </div>
+
+      <div className="card p-6 space-y-3 mb-6">
+        <h2 className="eyebrow">Campaign context</h2>
+        <textarea className="input w-full h-32" value={settings.world_context} onChange={(e) => set("world_context", e.target.value)} placeholder="Paste world context directly..." />
+        <input className="input w-full" value={settings.world_context_path} onChange={(e) => set("world_context_path", e.target.value)} placeholder="...or path to a file inside the container" />
+      </div>
+
+      <button className="btn-primary" onClick={save}>{saved ? "Saved" : "Save settings"}</button>
+    </div>
+  );
+}
--- a/frontend/src/pages/Setup.tsx
+++ b/frontend/src/pages/Setup.tsx
@@ -0,0 +1,205 @@
+import { useState, useEffect } from 'react';
+import { api, AppSettings } from '../api';
+
+const DEFAULT_SETTINGS: AppSettings = {
+  onboarding_completed: false,
+  whisper_model: 'medium',
+  whisper_compute_type: 'int8',
+  hf_token: '',
+  llm_mode: 'ollama',
+  ollama_host: 'http://localhost:11434',
+  ollama_model: 'qwen2.5:7b',
+  api_base_url: 'https://api.openai.com/v1',
+  api_key: '',
+  api_model: 'gpt-4o-mini',
+  chunk_word_target: '2500',
+  world_context: '',
+  world_context_path: '',
+};
+
+export default function Setup({ onComplete }: { onComplete: () => void }) {
+  const [step, setStep] = useState(1);
+  const [settings, setSettings] = useState<AppSettings>(DEFAULT_SETTINGS);
+
+  useEffect(() => {
+    api.getSettings().then(data => {
+      if (data) {
+        setSettings({ ...DEFAULT_SETTINGS, ...data });
+      }
+    }).catch(err => {
+      console.error('Failed to connect to settings endpoint:', err);
+    });
+  }, []);
+
+  const handleNext = () => setStep(prev => prev + 1);
+  const handleBack = () => setStep(prev => prev - 1);
+
+  const handleSave = async () => {
+    await api.saveSettings({ ...settings, onboarding_completed: true });
+    onComplete();
+  };
+
+  return (
+    <div className="min-h-screen bg-slate-900 text-slate-100 flex items-center justify-center p-6">
+      <div className="w-full max-w-2xl bg-slate-800 border border-slate-700 rounded-xl p-8 shadow-2xl">
+        <div className="mb-8 flex justify-between items-center">
+          <h1 className="text-2xl font-bold tracking-tight text-teal-400">Nat20Notes Configuration Wizard</h1>
+          <span className="text-sm text-slate-400 font-mono">Step {step} of 2</span>
+        </div>
+
+        {step === 1 && (
+          <div className="space-y-6">
+            <div>
+              <h2 className="text-xl font-semibold mb-2">1. Audio & Diarization Engine</h2>
+              <p className="text-slate-400 text-sm mb-4">Set up your local Whisper sizing and secure deep speaker parsing models.</p>
+            </div>
+
+            <div>
+              <label className="block text-sm font-medium mb-2">Whisper Model Size</label>
+              <select
+                value={settings.whisper_model}
+                onChange={e => setSettings({ ...settings, whisper_model: e.target.value })}
+                className="w-full bg-slate-900 border border-slate-700 rounded px-3 py-2 text-white"
+              >
+                <option value="tiny">Tiny (~39M params)</option>
+                <option value="base">Base (~74M params)</option>
+                <option value="small">Small (~244M params)</option>
+                <option value="medium">Medium (~769M params)</option>
+                <option value="large-v3">Large V3 (~1.5B params - Recommended GPU)</option>
+              </select>
+            </div>
+
+            <div className="bg-slate-900 p-4 rounded border border-slate-700 space-y-3">
+              <label className="block text-sm font-medium">Hugging Face Access Token</label>
+              <input
+                type="password"
+                placeholder="hf_..."
+                value={settings.hf_token}
+                onChange={e => setSettings({ ...settings, hf_token: e.target.value })}
+                className="w-full bg-slate-800 border border-slate-700 rounded px-3 py-2 text-white font-mono"
+              />
+              <div className="text-xs text-slate-400 leading-relaxed">
+                💡 <strong>Diarization Requirements:</strong> To differentiate between multiple speakers, you must register a token:
+                <ol className="list-decimal list-inside ml-2 mt-1 space-y-1">
+                  <li>Create an account on <a href="https://huggingface.co" target="_blank" rel="noreferrer" className="text-teal-400 underline">huggingface.co</a></li>
+                  <li>Accept licensing conditions for <a href="https://huggingface.co/pyannote/speaker-diarization-3.1" target="_blank" rel="noreferrer" className="text-teal-400 underline">pyannote/speaker-diarization-3.1</a></li>
+                  <li>Accept licensing conditions for <a href="https://huggingface.co/pyannote/segmentation-3.0" target="_blank" rel="noreferrer" className="text-teal-400 underline">pyannote/segmentation-3.0</a></li>
+                  <li>Generate a <strong>Read</strong> token in Settings → Access Tokens.</li>
+                </ol>
+              </div>
+            </div>
+
+            <div className="flex justify-end pt-4">
+              <button onClick={handleNext} className="bg-teal-600 hover:bg-teal-500 text-white px-5 py-2 rounded font-medium">
+                Continue Next
+              </button>
+            </div>
+          </div>
+        )}
+
+        {step === 2 && (
+          <div className="space-y-6">
+            <div>
+              <h2 className="text-xl font-semibold mb-2">2. Summarization Engine (LLM)</h2>
+              <p className="text-slate-400 text-sm mb-4">Choose where your post-transcription formatting and smart summaries are generated.</p>
+            </div>
+
+            <div className="flex gap-4 border-b border-slate-700 pb-4">
+              {(['ollama', 'api'] as const).map(mode => (
+                <button
+                  key={mode}
+                  onClick={() => setSettings({ ...settings, llm_mode: mode })}
+                  className={`px-4 py-2 rounded uppercase text-xs font-bold tracking-wider ${settings.llm_mode === mode ? 'bg-teal-600 text-white' : 'bg-slate-900 text-slate-400'}`}
+                >
+                  {mode === 'ollama' ? 'Local (Ollama)' : 'Hosted API'}
+                </button>
+              ))}
+            </div>
+
+            {settings.llm_mode === 'ollama' ? (
+              <div className="space-y-4">
+                <div>
+                  <label className="block text-sm font-medium mb-1">Ollama Host Address</label>
+                  <input
+                    type="text"
+                    value={settings.ollama_host}
+                    onChange={e => setSettings({ ...settings, ollama_host: e.target.value })}
+                    className="w-full bg-slate-900 border border-slate-700 rounded px-3 py-2"
+                  />
+                </div>
+                <div>
+                  <label className="block text-sm font-medium mb-1">Model Name</label>
+                  <input
+                    type="text"
+                    placeholder="qwen2.5:7b"
+                    value={settings.ollama_model}
+                    onChange={e => setSettings({ ...settings, ollama_model: e.target.value })}
+                    className="w-full bg-slate-900 border border-slate-700 rounded px-3 py-2 font-mono"
+                  />
+                  <p className="text-xs text-slate-500 mt-1">Must match your local CLI entry for running "ollama pull &lt;model&gt;".</p>
+                </div>
+              </div>
+            ) : (
+              <div className="space-y-4">
+                <div>
+                  <label className="block text-sm font-medium mb-1">API Base URL</label>
+                  <input
+                    type="text"
+                    value={settings.api_base_url}
+                    onChange={e => setSettings({ ...settings, api_base_url: e.target.value })}
+                    className="w-full bg-slate-900 border border-slate-700 rounded px-3 py-2 font-mono"
+                  />
+                </div>
+                <div>
+                  <label className="block text-sm font-medium mb-1">API Key</label>
+                  <input
+                    type="password"
+                    value={settings.api_key}
+                    onChange={e => setSettings({ ...settings, api_key: e.target.value })}
+                    className="w-full bg-slate-900 border border-slate-700 rounded px-3 py-2 font-mono text-white"
+                    placeholder="sk-..."
+                  />
+                </div>
+                <div>
+                  <label className="block text-sm font-medium mb-1">Model Name</label>
+                  <input
+                    type="text"
+                    value={settings.api_model}
+                    onChange={e => setSettings({ ...settings, api_model: e.target.value })}
+                    className="w-full bg-slate-900 border border-slate-700 rounded px-3 py-2 font-mono"
+                  />
+                </div>
+              </div>
+            )}
+
+            <div className="border-t border-slate-700 pt-4 space-y-3">
+              <h3 className="text-sm font-semibold text-slate-300">Campaign context (optional)</h3>
+              <textarea
+                className="w-full bg-slate-900 border border-slate-700 rounded px-3 py-2 h-24"
+                placeholder="Paste world/campaign context directly..."
+                value={settings.world_context}
+                onChange={e => setSettings({ ...settings, world_context: e.target.value })}
+              />
+              <input
+                type="text"
+                className="w-full bg-slate-900 border border-slate-700 rounded px-3 py-2 font-mono text-sm"
+                placeholder="...or point to a local file"
+                value={settings.world_context_path}
+                onChange={e => setSettings({ ...settings, world_context_path: e.target.value })}
+              />
+            </div>
+
+            <div className="flex justify-between pt-4 border-t border-slate-700">
+              <button onClick={handleBack} className="bg-slate-700 hover:bg-slate-600 text-white px-5 py-2 rounded font-medium">
+                Back
+              </button>
+              <button onClick={handleSave} className="bg-emerald-600 hover:bg-emerald-500 text-white px-5 py-2 rounded font-medium">
+                Complete Setup & Launch
+              </button>
+            </div>
+          </div>
+        )}
+      </div>
+    </div>
+  );
+}
--- a/frontend/src/pages/Speakers.tsx
+++ b/frontend/src/pages/Speakers.tsx
@@ -0,0 +1,181 @@
+import { useEffect, useRef, useState } from "react";
+import { Link, useNavigate, useParams } from "react-router-dom";
+import { api } from "../api";
+import { SessionReel, colorFor } from "../components/SessionReel";
+
+export default function Speakers() {
+  const { id } = useParams();
+  const nav = useNavigate();
+  const [speakers, setSpeakers] = useState<any[]>([]);
+  const [turns, setTurns] = useState<any[]>([]);
+  const [names, setNames] = useState<Record<string, string>>({});
+  const [currentTime, setCurrentTime] = useState(0);
+  const [duration, setDuration] = useState(0);
+  const [selected, setSelected] = useState<Set<string>>(new Set());
+  const [mergeTarget, setMergeTarget] = useState("");
+  const [busy, setBusy] = useState(false);
+  const [showConfirm, setShowConfirm] = useState(false);
+  const audioRef = useRef<HTMLAudioElement>(null);
+  const paletteRef = useRef(new Map<string, string>());
+
+  const load = () => {
+    api.getSpeakers(id!).then((sp) => {
+      setSpeakers(sp);
+      const n: Record<string, string> = {};
+      sp.forEach((s: any) => (n[s.raw_label] = s.display_name || ""));
+      setNames(n);
+    });
+    api.getTurns(id!).then(setTurns);
+  };
+  useEffect(load, [id]);
+
+  const save = async (raw_label: string) => {
+    await api.setSpeakerName(id!, raw_label, names[raw_label] || "");
+  };
+
+  const playAt = (t: number) => {
+    const audio = audioRef.current;
+    if (!audio) return;
+    if (audio.readyState < 2) return;
+    audio.pause();
+    audio.currentTime = t;
+    const p = audio.play();
+    if (p) p.catch(() => {});
+  };
+
+  const toggleSelected = (label: string) => {
+    const next = new Set(selected);
+    if (next.has(label)) next.delete(label); else next.add(label);
+    setSelected(next);
+  };
+
+  const doMerge = async () => {
+    if (!mergeTarget || selected.size < 2) return;
+    setBusy(true);
+    const sourceLabels = [...selected].filter(l => l !== mergeTarget);
+    await fetch(`/api/sessions/${id}/speakers/merge`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ target_label: mergeTarget, source_labels: sourceLabels }),
+    });
+    setSelected(new Set());
+    setMergeTarget("");
+    setBusy(false);
+    load();
+  };
+
+  const handleDone = async () => {
+    setBusy(true);
+    await api.generateNotes(id!);
+    setBusy(false);
+    nav(`/sessions/${id}`);
+  };
+
+  const selectedArr = [...selected];
+
+  return (
+    <div className="max-w-4xl mx-auto py-16 px-4">
+      <Link to={`/sessions/${id}`} className="text-sm text-ink/40 hover:text-brass">&larr; Back</Link>
+      <h1 className="font-display text-3xl mt-4 mb-2">Name your speakers</h1>
+      <p className="text-ink/60 mb-6">Click anywhere on the reel to jump to that moment and hear who's talking.</p>
+
+      <audio
+        ref={audioRef}
+        src={api.audioUrl(id!)}
+        onTimeUpdate={(e) => setCurrentTime(e.currentTarget.currentTime)}
+        onLoadedMetadata={(e) => setDuration(e.currentTarget.duration)}
+        controls
+        className="w-full mb-3"
+      />
+
+      <div className="mb-8">
+        <SessionReel turns={turns} duration={duration} currentTime={currentTime} onSeek={playAt} />
+      </div>
+
+      {selected.size >= 2 && (
+        <div className="card p-4 mb-6 flex items-center gap-3 flex-wrap">
+          <span className="text-sm text-ink/60">{selected.size} selected</span>
+          <select
+            className="input text-sm flex-1 min-w-[200px]"
+            value={mergeTarget}
+            onChange={(e) => setMergeTarget(e.target.value)}
+          >
+            <option value="">Merge into...</option>
+            {selectedArr.map((l) => (
+              <option key={l} value={l}>{l} {names[l] ? `(${names[l]})` : ""}</option>
+            ))}
+          </select>
+          <button className="btn-primary text-sm" disabled={!mergeTarget || busy} onClick={doMerge}>
+            {busy ? "Merging..." : "Merge"}
+          </button>
+          <button className="btn-secondary text-sm" onClick={() => { setSelected(new Set()); setMergeTarget(""); }}>
+            Cancel
+          </button>
+        </div>
+      )}
+
+      <div className="grid gap-4 sm:grid-cols-2">
+        {speakers.map((sp) => (
+          <div key={sp.raw_label} className={"card p-4 " + (selected.has(sp.raw_label) ? "ring-2 ring-brass" : "")}>
+            <div className="flex items-center gap-2 mb-2">
+              <input
+                type="checkbox"
+                checked={selected.has(sp.raw_label)}
+                onChange={() => toggleSelected(sp.raw_label)}
+                className="accent-brass"
+              />
+              <span
+                className="w-3 h-3 rounded-full inline-block flex-shrink-0"
+                style={{ background: colorFor(sp.raw_label, paletteRef.current) }}
+              />
+              <span className="font-mono text-xs text-ink/40">{sp.raw_label}</span>
+            </div>
+            <input
+              className="input w-full mb-3"
+              placeholder="Player or character name"
+              value={names[sp.raw_label] || ""}
+              onChange={(e) => setNames({ ...names, [sp.raw_label]: e.target.value })}
+              onBlur={() => save(sp.raw_label)}
+            />
+            <div className="space-y-1">
+              {sp.samples.map((s: any, i: number) => (
+                <button
+                  key={i}
+                  onClick={() => playAt(s.start)}
+                  className="block text-left text-sm text-ink/60 hover:text-ink w-full truncate"
+                >
+                  ▶ {s.text}
+                </button>
+              ))}
+            </div>
+          </div>
+        ))}
+      </div>
+
+      <div className="mt-8">
+        <button className="btn-primary" onClick={() => setShowConfirm(true)}>
+          Done naming
+        </button>
+      </div>
+
+      {showConfirm && (
+        <div className="fixed inset-0 bg-black/60 flex items-center justify-center p-4 z-50">
+          <div className="bg-panel border border-white/10 max-w-md w-full rounded-xl p-6 shadow-2xl">
+            <h4 className="text-lg font-bold mb-2">Generate notes now?</h4>
+            <p className="text-sm text-ink/60 mb-6 leading-relaxed">
+              This will use your speaker names to create session notes. Are you sure?
+            </p>
+            <div className="flex gap-3 justify-end">
+              <button onClick={() => setShowConfirm(false)} className="btn-secondary">
+                Cancel
+              </button>
+              <button disabled={busy} onClick={handleDone} className="btn-primary">
+                {busy ? "Starting..." : "Yes, generate notes"}
+              </button>
+            </div>
+          </div>
+        </div>
+      )}
+    </div>
+  );
+}
--- a/frontend/src/styles.css
+++ b/frontend/src/styles.css
@@ -0,0 +1,32 @@
+@tailwind base;
+@tailwind components;
+@tailwind utilities;
+
+@layer base {
+  ::selection {
+    background: #C9A227;
+    color: #14171F;
+  }
+  :focus-visible {
+    outline: 2px solid #C9A227;
+    outline-offset: 2px;
+  }
+}
+
+@layer components {
+  .card {
+    @apply bg-panel border border-white/5 rounded-lg;
+  }
+  .btn-primary {
+    @apply bg-brass text-deep font-medium px-4 py-2 rounded-md hover:brightness-110 transition disabled:opacity-40 disabled:pointer-events-none;
+  }
+  .btn-secondary {
+    @apply bg-panel2 text-ink font-medium px-4 py-2 rounded-md border border-white/10 hover:border-brass/50 transition;
+  }
+  .input {
+    @apply bg-deep border border-white/10 rounded-md px-3 py-2 text-ink placeholder:text-ink/30 focus:border-brass/60;
+  }
+  .eyebrow {
+    @apply font-mono text-xs uppercase tracking-widest text-brass/80;
+  }
+}
--- a/frontend/tailwind.config.js
+++ b/frontend/tailwind.config.js
@@ -0,0 +1,26 @@
+/** @type {import('tailwindcss').Config} */
+export default {
+  content: ["./index.html", "./src/**/*.{ts,tsx}"],
+  theme: {
+    extend: {
+      colors: {
+        deep: "#14171F",
+        panel: "#1E2230",
+        panel2: "#262B3D",
+        ink: "#EDE6D6",
+        inkdim: "#A9A4956e".slice(0, 7),
+        brass: "#C9A227",
+        ember: "#B4523A",
+        moss: "#5F7A61",
+        dusk: "#4E6A87",
+        plum: "#7A5670",
+      },
+      fontFamily: {
+        display: ["Fraunces", "Georgia", "serif"],
+        body: ["Inter", "system-ui", "sans-serif"],
+        mono: ["IBM Plex Mono", "monospace"],
+      },
+    },
+  },
+  plugins: [],
+};
--- a/frontend/tsconfig.json
+++ b/frontend/tsconfig.json
@@ -0,0 +1,16 @@
+{
+  "compilerOptions": {
+    "target": "ES2020",
+    "useDefineForClassFields": true,
+    "lib": ["ES2020", "DOM", "DOM.Iterable"],
+    "module": "ESNext",
+    "skipLibCheck": true,
+    "moduleResolution": "bundler",
+    "resolveJsonModule": true,
+    "isolatedModules": true,
+    "noEmit": true,
+    "jsx": "react-jsx",
+    "strict": true
+  },
+  "include": ["src"]
+}
--- a/frontend/vite.config.ts
+++ b/frontend/vite.config.ts
@@ -0,0 +1,11 @@
+import { defineConfig } from "vite";
+import react from "@vitejs/plugin-react";
+
+export default defineConfig({
+  plugins: [react()],
+  server: {
+    proxy: {
+      "/api": "http://localhost:8000",
+    },
+  },
+});