Case study / AI product
Building an Autonomous AI Video Production Platform
A look at the orchestration and reliability choices behind an AI video production platform.
- Product
- AI video production platform
- Role
- Founding Engineer & AI Systems Architect
- Period
- 2026–present
- Status
- Active product
- Stack
- TypeScript, React, Fastify, PostgreSQL, Redis, BullMQ, FFmpeg
The problem
Turning a creative idea into a finished video requires planning, scene and asset generation, inspection, assembly and rendering. These stages depend on external model providers, take different amounts of time and may fail independently. A synchronous request cannot safely own the whole workflow.
My contribution
I designed a TypeScript modular monolith with separate API, SSE, worker and scheduler processes. BullMQ and Redis coordinate dedicated worker pools, while CPU-heavy FFmpeg rendering runs apart from the API. This gives each kind of work its own resource and failure boundary without splitting the product into many services.
I built the React 19 studio interface with a multi-track timeline and streaming AI Director panel. On the backend I implemented a Fastify API, PostgreSQL data model, authentication and rate limiting, immutable version rows and a transactional double-entry credit ledger.
AI workflow and state
The AI Director proposes structured commands for scenes and shots. The backend validates each proposal against budgets, locks and schemas before it changes project state. The model does not write directly to the database or storage.
Provider adapters connect to Gemini, Imagen, Veo, ElevenLabs, OpenRouter and Claude. Backoff, circuit breaking and content-hash caching help with provider failures and repeated work. Redis Pub/Sub and Server-Sent Events deliver progress with monotonic IDs so clients can reconnect using Last-Event-ID.
Reliability decisions
Long-running render jobs are isolated so they cannot stall normal API requests. Append-only project versions make previous states recoverable. PostgreSQL row locks protect concurrent credit deductions. These are architecture decisions aimed at predictable behavior under retries, concurrency and partial failure.
Outcome and evidence
The resulting product connects planning, generation, inspection, editing and rendering in one workflow. I do not publish unverified throughput or revenue metrics. The public product site provides further context.
Related services
AI Agent Development · SaaS Development · Full-Stack Development · Backend & API Development
Have a project in mind?
Let’s talk about an AI product.
Share your goals, current system and timeline. I can help scope the architecture and implementation.