Sri Studio

AI reel generator. A text brief becomes a vertical short-form video with voice-cloned narration, image scenes, and burn-in captions, with a mandatory human-in-the-loop approval gate before any paid media APIs fire.

Live: https://studio.sshub.dev/ (basic auth; credentials shared in the submission email)

What it does

1. Extract the brief into structured intent (brand, audience, duration, mood) via Claude Haiku.
2. Plan a ScriptPlan (hook, voiceover text, per-scene visual prompts, motion) via Claude Sonnet.
3. Approve / edit the plan in the browser. The user can rewrite the hook, voiceover, and per-scene prompts before anything bills. Edits are captured to runs/<id>/plan_edits.json for prompt distillation.
4. Execute in parallel: Flux Schnell for 9:16 images, ElevenLabs for cloned-voice TTS plus alignment-driven captions.
5. Stitch with ffmpeg into a 1080x1920 MP4.

Pipeline is a LangGraph state machine with deterministic node order, traced end to end on Langfuse.

Repository layout

backend/        FastAPI + LangGraph pipeline, pytest suite (24 tests)
frontend/       Next.js 14 app (typed, Tailwind), runs page + plan-review UI
docs/
  decisions/    ADRs (LangGraph over n8n, Flux over SDXL, runs-as-filesystem, ...)
  superpowers/  Spec and 11-phase implementation plan
deploy/         tar-over-ssh deploy script for Hetzner
nginx/          studio.sshub.dev system-nginx site config
samples/        End-to-end sample reels (mp4 + plan.json + cost.json)
runs/           Per-run artifacts (gitignored)

Local dev

cp .env.example .env       # fill in OpenRouter, Langfuse, Replicate, ElevenLabs keys
docker compose up --build  # backend on :8000, frontend on :3000

Open http://localhost:3000 and submit a brief.

Backend tests

cd backend
uv pip install -e ".[dev]"
uv run pytest

Frontend typecheck

cd frontend
npm install
npm run typecheck

Cost discipline

Every LLM and media call writes to runs/<id>/cost.json. A daily cap (DAILY_COST_CAP_USD, default $5) is enforced server-side in runs/_daily_cost.json. The approval gate is the primary cost control: if the plan looks wrong, the user rejects before any image, TTS, or stitch call fires. Sample reels in samples/ cost $0.03 to $0.06 each end to end. A rejected run cost $0.0058 (Extract + Plan only).

Deployment

The production host runs Docker Compose behind system nginx + Let's Encrypt + basic auth. Both containers bind to 127.0.0.1 only; the host's nginx routes studio.sshub.dev/api/* to the backend container and studio.sshub.dev/ to the Next.js container. See nginx/studio.sshub.dev.conf and deploy/deploy.sh.

Design choices worth knowing

• Human approval is mandatory, not optional. ADR 012. The gate exists because text-to-video pipelines are easy to drive into expensive runaway loops.
• Runs are filesystem-backed, no database. ADR 010. Each run is runs/<id>/ with intent.json, plan.json, plan_edits.json, cost.json, scene images, audio, captions, and the final reel.mp4. Trivially diffable, archivable, and shippable as samples.
• LangGraph over n8n. ADR 001. Typed state, version-controlled graph, native Python testability.
• 5-second default with 90 seconds on tape. ADR 004. The model plans long, the stitcher trims short, the user always has the source.
• Plan edits are training signal. Every edit at the approval gate is captured with original-vs-edited diff. Future work: aggregate edits to identify systemic Plan-prompt failures and distill better defaults.

Sample reels

See samples/README.md. Three end-to-end runs from the live system: a Sri Studio launch teaser, a Sur La Table spring sale, and a deliberately rejected Cozy autumn coffee shop brief.