Field Repair Assistant
Installable mobile web app that guides an appliance technician through a repair: guided photos, AI reading of the rating plate, a branching step-by-step plan, evidence review, quote and customer report.
Architecture
Next.js App Router with PostgreSQL and Drizzle. Vision models extract model and serial data from photos as Zod-validated structured outputs, and the technician confirms before anything is planned; reviewing evidence can re-plan the rest of the job. Photos upload to S3 through presigned URLs, and an IndexedDB outbox holds work captured without signal. Runs in Docker on EC2 behind CloudFront; the origin only accepts CloudFront traffic carrying a secret header, and secrets live in SSM Parameter Store.
Highlights
- Live in production
- Offline-first photo capture and sync
- Human confirmation before every AI step
- Tiered model escalation with timeouts under the CDN limit
- Classifier evals and a zero-cost mock mode