Scaffold — Compound AI for Product Photography

Vercel
application badge
Created by team Scaffold on July 06, 2026
Unicorn Track

Scaffold is a compound AI inference router that powers ProductShot, a product-photography tool. Instead of sending every request to one expensive model, Scaffold reads each task and routes it to the cheapest model that can still deliver — a low-cost compose path (background removal plus generation) for simple shots, or a higher-quality edit path for complex ones. Routing intelligence is driven by a lightweight model that classifies each scene, while a stronger verifier checks the output and triggers a self-healing retry when quality falls short. Every routing decision is logged to a telemetry store and surfaced on a live, public dashboard, so the system's decisions are fully transparent and auditable in real time. The result is a system that holds output quality while cutting cost per image dramatically, because expensive inference is reserved only for the tasks that actually need it. Judges can operate the product end-to-end through a live demo — generating a real image without signing up — and inspect the routing dashboard directly. Scaffold is built and deployed on modern serverless infrastructure with Fireworks AI for model inference, demonstrating a practical, product-ready approach to compound AI: real routing, real verification, real telemetry, and a working live demo rather than a benchmark run.

Category tags: