← All work
AI systems, 2024 – 2026

ACT3 AI — filmmaking platform

Script to storyboard to cinematography to render in one web app. Two years as a top-three contributor on the client and the API.

Role
Senior engineer, top-3 contributor on both main repos
Context
Whitehat Engineering for ACT3 AI
Scale
~1,750 client commits, ~950 API commits, ~4,000 source files, team of ~10

Background

ACT3 AI is a professional tool for making animated shows with generative models. Writers edit scripts, the system proposes storyboards and shot compositions, and a render pipeline produces frames and video. It is multi-tenant with organisations, credits and Stripe billing.

I have worked across the whole product since 2024: the Next.js client, the NestJS API, a FastAPI service for script generation and RAG, and the Python Blender automation for shots and upscaling.

The problem

Generative video is non-deterministic and expensive. The product has to let a writer iterate on a shot dozens of times while keeping characters consistent, keeping cost visible, and keeping the UI responsive while a generation runs for minutes in the background.

Approach

  1. Client: built the new home surface with design tokens, frame drill-down and a rendered-frame stage, so a writer sees the episode as frames rather than as a job table.
  2. Seedance prompt tiers end to end: tier selection in the client, prompt families and generation queues in the API, so cheaper tiers are the default and expensive ones are a deliberate choice.
  3. Storyboard first-frame cast-reference resolver, so each shot's generation is seeded with the right character references and faces stay consistent across a scene.
  4. Script and scene generation in the FastAPI service on GPT-4o, o3 and Grok, with formal and informal screenplay import and a copyright filter.
  5. Procedural Blender shot and camera-path generation, Real-ESRGAN upscaling and episode assembly in the Python visual-automation pipeline.
  6. BullMQ generation queues with WebSocket progress, a content-safety layer and Sentry across client, API and services.

Architecture

showrunnerhero-appclientNext.js 16 + MUI, generated API hooks, ~2,400 files.
aitaskassembly-appsrvNestJS 9, PostGIS + pgvector, BullMQ, Socket.IO, Stripe, SendGrid, AWS.
ai-serviceFastAPI, pgvector RAG, script and scene generation.
visual-automationPython, Blender bpy, Real-ESRGAN, YOLO. Second-largest contributor.
job_runnerTemporal job orchestration. See its own case study.

Outcome

  • Production platform with paying organisations, in continuous development since 2024.
  • Third-largest contributor on the client and fourth on the API over two years, across a team of about ten engineers.

Lessons

  • On a large team codebase, the highest-leverage work is the seam between systems: the tier pipeline touched client, API and runner, and that is where the product value was.
  • Generated API clients from OpenAPI pay for themselves in a week on a team this size.

Stack

Next.jsReactMUINestJSPostgreSQLpgvectorTypeORMBullMQSocket.IOStripeLangChainAWSFastAPIBlender