WW2 Headquarters AI Lab

docker-compose.yml — Stack Services

services: db: pgvector/pgvector:pg16 # PostgreSQL + vector search redis: redis:7-alpine # Job queue broker api: ./api # FastAPI — 0.0.0.0:8000 worker: ./api (celery) # Background agent worker dashboard: ./dashboard # Next.js — 0.0.0.0:3000 volumes: pg_data · redis_data · media_store (/media)

Component Stack

Orchestration

Docker Compose

Single command brings up the full stack — API, worker, dashboard, DB, Redis

Backend API

FastAPI (Python 3.12)

REST API for all pipeline operations — ingest, review, approve, publish

Job Queue

Celery + Redis

Long-running AI jobs run in background — transcription can take 10–20 min

Database

PostgreSQL 16 + pgvector

Stores all jobs, transcripts, clips, captions. pgvector for Knowledge DB semantic search

Downloader

yt-dlp

Pulls video + metadata from YouTube in best available quality (≤720p)

Transcription

OpenAI Whisper (local)

Runs on-VPS — no data leaves the machine. Timestamped transcript with segments

AI Intelligence

Anthropic Claude API

Selects best clip moments, writes platform captions, analyses performance

Dashboard

Next.js 15

Human review UI — approve/reject clips, edit captions, monitor pipeline status

VPS Access

Tailscale

Zero-config private network. Chris and Peter access the dashboard securely. Public demo URL via Funnel

Video Edit

DaVinci / Premiere API

Phase 3 — AI triggers rough cut assembly via scripting APIs

Pipeline — URL In, Captions Out

Submit YouTube URL

Dashboard → POST /ingest

Download video + metadata

yt-dlp → mp4 + JSON

Generate transcript

OpenAI Whisper (local, on-VPS)

Identify best clips

Claude API → JSON with timestamps

Cut clips

ffmpeg → clip_1.mp4, clip_2.mp4...

Write captions

Claude API → Instagram, YouTube, TikTok, X

Human review

Dashboard → approve / edit / reject

Load into scheduler

Phase 2 — social publishing queue

Known Technical Challenges

Grading of visual — colour consistency

Drone, camera and archive footage have different colour profiles. Phase 3 uses DaVinci colour node scripting to normalise; requires approved LUT from Jim’s editor.

Sound sync between different shots

Multi-source audio (lavalier, boom, room) requires alignment. ffmpeg auto-sync works for single-cam; multi-cam sync needs DaVinci scripting in Phase 3.

Access Model

VPS — DigitalOcean LON1

Spec

4 vCPU / 8 GB RAM + 100 GB volume

Ubuntu 22.04 LTS

Access

Tailscale — private network only

Dashboard

http://jhwwii-lab:3000 (Tailscale hostname)

API docs

http://jhwwii-lab:8000/docs

Public demo

tailscale funnel --bg 3000 → HTTPS URL

CI access

Chris + Peter added to Tailscale network

The engine.For Peter’s team.

The engine.
For Peter’s team.