Reimagining Urgent Shopping: From intent to confirmed order in under 30 seconds.
Built for Amazon HackOn 2026 Theme: Amazon Now Reimagining Urgent Shopping
Quick-commerce delivers in 10 minutes. But the customer spends 15 minutes deciding what to order. The bottleneck isn't the rider, it's the shopping itself.
| # | Feature | What It Does |
|---|---|---|
| 1 | Intent-to-Cart Compiler | 🎤 Speak "make paneer butter masala" → LLM decomposes → vector search → full purchasable cart in 5 seconds |
| 2 | Probabilistic Substitution Integrity Layer | Out-of-stock item → auto-suggests closest alternative (60% semantic + 30% price + 10% brand scoring) |
| 3 | Autonomous Cart Stability System | Cart self-corrects on price changes, stock fluctuations, and network drops. Event-driven + offline-first. |
| 4 | Consumption Forecasting & Replenishment Engine | Learns purchase rhythm from 2nd order. Overdue items surface automatically. Zero setup. |
| 5 | Anti-Double Spend Inventory Enforcement Layer | Redis atomic DECRBY ensures each unit reserved exactly once. No overselling under any concurrency. |
| Layer | Technology |
|---|---|
| Frontend | React 19, Vite 8, Redux Toolkit, Tailwind CSS 4, Web Speech API |
| Backend | Go (Gin), JWT Auth, RabbitMQ consumers |
| ML Engine | Python FastAPI, sentence-transformers (MiniLM-L12-v2) |
| AI | Google Gemini (task decomposition) |
| Vector DB | PostgreSQL + pgvector (384-dim embeddings, cosine similarity) |
| Operational DB | MongoDB Atlas |
| Cache | Redis (inventory counters, stale flags, replenishment stats) |
| Event Bus | RabbitMQ (topic exchange: price.changed, stock.changed, user.purchase) |
| Infra | AWS EC2, Docker Compose |
- Docker Desktop
- Go 1.25+
- Node.js 22+
- Python 3.10+
make buildThis installs all dependencies (Go, Node, Python/PyTorch), starts Docker infra, and seeds the ML engine with product embeddings.
make run| Service | URL |
|---|---|
| Frontend | http://localhost:5173 |
| Backend API | http://localhost:8080 |
| ML Engine | http://localhost:8000 |
| RabbitMQ UI | http://localhost:15672 |
| Admin Panel | http://localhost:3001 |
make stop├── Client/ # React frontend (Vite + Redux + Tailwind)
│ └── src/
│ ├── components/ # UI (Header, Cart, Checkout, Search, TaskShopping)
│ └── features/ # Redux slices (cart, auth, orders, recommendations, stability)
│
├── Server/ # Go backend (Gin)
│ ├── controllers/ # Auth, Cart, Products, Orders, Recommendations, Admin
│ ├── services/ # Redis, RabbitMQ, LLM, Inventory, Consumers
│ ├── middleware/ # JWT Auth, CORS
│ └── models/ # MongoDB document schemas
│
├── ml_engine/ # Python FastAPI (sentence-transformers + pgvector)
│ └── main.py # /search, /find-substitute, /seed endpoints
│
├── admin/ # Standalone admin panel (port 3001)
├── docker-compose.yml # Redis + RabbitMQ + PostgreSQL+pgvector
└── Makefile # make build | make run | make stop
- Hybrid Vector + Keyword Search: Cosine similarity (pgvector) merged with ILIKE keyword boost. O(log n) with IVFFlat index.
- Probabilistic Substitution Scoring:
0.6×semantic + 0.3×price_proximity + 0.1×brand_match - Replenishment Formula:
weighted_score = 0.6 × (gap/avg_interval) + 0.4 × log(count+1). Works from 2nd purchase, no ML training. - Atomic Inventory: Redis DECRBY (O(1), sub-ms). Same pattern as ticketing systems handling 100k concurrent requests.
Team Duo: Built for Amazon HackOn 2026