AI video generation platform

The situation

Generative video moves fast. Models appear, change price and get deprecated within months, and a platform built around any single one of them has a short life.

What we built

• Nineteen video models behind a single internal API, each with its own capabilities, parameters and per-second pricing
• A model roster that handles deprecation, so retired models do not break saved projects
• A look system that keeps characters, lighting and grade consistent across a sequence of shots
• Storyboarding, lipsync and upscaling as first-class steps rather than separate tools
• A credit ledger with per-generation cost accounting that reconciles against provider invoices
• Job queues with retries, progress reporting and dead-letter handling
• An MCP server so AI agents can drive the platform directly

Engineering notes

Per-second pricing across nineteen models with different cost structures is the part that looks trivial and is not. Every model prices differently — by duration, resolution, frame rate or step count — and the ledger has to predict cost before a job runs, hold the credits, and reconcile afterwards against what the provider actually charged.