Improve field fidelity, latency and retry QA in a pool-design AI pipeline
Project brief
Optimize an existing pool-design generator using Next.js, TypeScript, Prisma and Supabase. It uses Meta AI for sketches, OpenAI for renders and Claude for QA; a cycle currently takes about 3-5 minutes.
- Keep intake dimensions, materials, coping and colors faithful in prompts; remove invented details. - Profile chained cleanup/pool-addition calls and QA, then make safe latency improvements with measured before/after timings. - Tighten field-level QA, including no-photo site disclosure, actionable corrections and the existing two-attempt retry accounting.
Deliver revised templates/code, comparison outputs, benchmarks and QA notes. Confirm feasibility and priorities on day one of the production week. Authorized access, fixtures and API allowance must be agreed first.
Deliverables & acceptance
What you'll deliver
- Updated pipeline, before/after evidence and QA documentation.
What the result must meet
- Tests demonstrate field fidelity, measured timings, specific QA feedback and correct retry consumption/reset.