Scaffold — Compound AI for Product Photography

Scaffold is a compound AI system powering ProductShot: sellers turn one phone photo into professional product photography. Instead of calling one image model per request, Scaffold routes each generation through the cheapest architecture that meets the quality bar, and a verifier watches every output, self-healing when needed. The problem. AI photo tools call FLUX Kontext per image at ~$0.04 — justified for complex scenes, overkill for simple ones (studio white, flatlay, moody dark). Kontext occasionally distorts brand-critical details like buckles, logos, collars, and sellers ship distorted images without knowing. Scaffold's answer. A per-scene router picks edit (FLUX Kontext Pro) for complex scenes, or compose (rembg + FLUX Schnell + sharp) for simple scenes needing preservation. Compose is 20x cheaper; preservation is guaranteed by construction — only the background is AI-generated, product pixels untouched. Compose is slightly slower (10.6s vs 9.8s), running three steps versus edit's one. Every generation is verified by Kimi K2.6 on Fireworks via structured JSON: similarity score, reasoning, distortions flagged. Below threshold, the router retries on compose — the self-heal. Every decision, cost, latency, and score logs to Supabase and renders on a live dashboard at /routing. Early data (N=23, growing): compose cost $0.006 vs edit $0.04, latency 10.6s vs 9.8s, Kimi score 0.92 vs 0.90, one self-heal captured 0.13ms apart — failed edit and successful retry, side by side. AMD integration. Kimi verification runs on Fireworks' platform on AMD MI300X — every self-heal call is AMD-accelerated inference in production today. Day-30: a dedicated preservation verifier on AMD Developer Cloud. Why this matters. AI photo tools compete on backdrop quality. Scaffold competes on reliability. ProductShot is shipped and monetized with paying users — Scaffold makes it 40% cheaper, guarantees preservation on 5 of 8 styles, self-heals when AI fails.