A camera-first app plus API that turns a barcode scan or a plain photo into a clean, reviewed product record and maintains stock from there: several product databases and a web search are queried in parallel, a model merges the candidates into one draft, the user confirms it, and where no barcode is readable, vision takes over. The wizard shows each resolution stage and the running AI cost live.
What ports as-is
- Identifier checksum validation before any lookup
- Parallel fan-out over pluggable data-source connectors
- An LLM merge into one normalized draft
- Server-side barcode detection on uploaded photos with preprocessing variants
- Vision fallback cached by normalized product name
- The product image pipeline (WebP, thumbnails, generated category artwork)
- Stock items with categories and quantity stepping
- Per-stage tracing with token and image cost
- An NDJSON stream pushing stage and cost to the UI
- Mixed directory and local auth with admin-gated user management
- Live camera scanning with pause/resume and a staged iOS constraint fallback
- A review-gated spreadsheet import for bulk master data
What we build for you
- Which data sources apply to the vertical (the connector set is the swap point)
- The target record schema and which fields the review step exposes
- The category taxonomy and custom attributes
- Branding
- Model choice and cost budget
- Tenant and redirect URIs
- One import mapping per spreadsheet the office already keeps
