Auto-Parts Warehouse Management
Quản lý Kho Phụ tùng Ô tô
Production WMS for a 14-branch business — barcode/QR, AI-vision stocktake, per-shift accounting reconciliation.
Problem
A parts business has 14 branches, 38,000 SKUs. Manual stocktake took 4 days/month × 5 people. Month-end variance was 3-5%. Shelves are 3-4 m tall — easy to miscount layers.
Architecture
Next.js 15 web (operator) + React Native scanner (mobile) → FastAPI backend → Postgres (multi-tenant via schemas) → offline-first IndexedDB queue on mobile. AI stocktake: photo a shelf → YOLOv8 detect & count → object tracking per layer → reconcile DB → flag mismatches.
Stack & rationale
- Postgres schema multi-tenant: simpler than DB-per-tenant, isolates 14 branches well.
- IndexedDB queue: scanner works fully offline; syncs when wifi returns.
- YOLOv8 + ByteTrack: more accurate counts than detect-only when items overlap.
Results (6 months)
| Metric | Before | After |
|---|---|---|
| Stocktake duration | 4 days | 0.5 days |
| Month-end variance | 3-5% | <0.3% |
| Outbound throughput | 1× | 2.4× |
| Scanner training needed | 5 people | 1 (UX is dead-simple) |