Senior Full-Stack Engineer — AI Video Production Pipeline (Next.js + Postgres + FFmpeg)

We've built most of an AI video production system. We need someone to finish it and make it genuinely excellent. We run a portfolio of faceless and avatar-led YouTube channels for a 60+ audience. We've built an internal tool meant to take a video from idea to finished file: script writing, research, shot planning, footage sourcing, AI avatar rendering, voiceover, word-level transcript alignment, and a final cut assembled with FFmpeg on our own server. Being straight with you about where it stands: the architecture is there and the code is written, but it has not yet produced a finished video end to end. Several vendor integrations were written against API documentation and have never been run against the live service. Getting the first real render out — and fixing everything that surfaces when we do — is a significant part of this job, not a formality. What you'd be working on First: make it actually run. Prove each integration against the real API, get a complete video out of the pipeline, and fix what that exposes. Expect surprises. Then: make the output good. - The assembler — an FFmpeg filter-graph builder in Node. Ken Burns on stills, push on footage, film-style dissolves, per-channel colour treatments, split-screen avatar layouts, drawtext title cards, loudness normalisation. It needs a better eye than it currently has. - Cut timing — we transcribe the voiceover with whisper.cpp and align shots to word-level timestamps so cuts land on the breath between words. Making the *pacing* feel deliberate rather than mechanical is open work. - Shot planning — an LLM writes the shot list against a per-channel "visual world" description. Better prompts, better structure, better judgement about when a still beats a clip. - Footage sourcing — stock (Pexels), public-domain archive film, and AI generation, chosen per shot. Relevance is mediocre and we know it. - Motion graphics — currently a title card. We want on-screen points, animated text, 2D/3D explainer visuals for educational channels. The stack Next.js (App Router), Postgres via `pg` with parameterised SQL, Docker Compose on a VPS, GitHub Actions deploys, Cloudflare in front. FFmpeg and whisper.cpp compiled into the image. A background job queue with claim-first semantics, per-vendor circuit breakers, and hard spend caps. Multiple AI vendors behind a capability registry (video, image, TTS, avatar, text) so no single provider is load-bearing. Roughly 70 database migrations and ~1,600 tests. Heavily commented — with *why*, not *what*. What we actually care about You need to be good at video, not just at code. The hardest problems here are "why does this cut feel wrong" and "why does this look AI-generated" — not "how do I write a React component". If you've done post-production or serious editing work alongside engineering, lead with that. You need to be careful with money. Every render costs real credits, and an untested integration in a retry loop can spend a lot before anyone notices. The codebase is built around that — charge before you spend, refuse rather than guess, make failures loud. You'd be expected to keep working that way. You should be comfortable reading before writing. There's a lot of accumulated reasoning to absorb, and a lot of decisions that look arbitrary until you read why. AI-assisted development is fine and expected — this was largely built that way, and it got us a long way. It has not got us to a finished, high-quality video, which is exactly why this post exists. To apply Skip the generic pitch. Tell me: 1. A video thing you've built or improved, ideally involving FFmpeg or an edit pipeline. What was wrong, what you changed, how you knew it was better. 2. The last time you shipped something that spent real money per call. How did you stop it running away? 3. Your honest read: what usually makes AI-generated video look AI-generated, and what you'd do about it. One paragraph each is plenty. I'd rather read three specific answers than a CV.