Indian quick-commerce platforms deliver groceries in 10 minutes, but catalog fragmentation and dynamic pricing mean that identical grocery baskets often differ in price by 20% or more.
To solve this, I built BasketBuddy—a React Native app that compares grocery baskets across stores in real time, auto-matches catalog entities, and bridges carts directly to store checkouts.
Here is an architectural breakdown of how it works and the engineering hurdles faced.
1. The Core Engineering Challenge: Catalog Entity Resolution
Store catalogs rarely share standardized SKUs or naming conventions:
- Store A: "Amul Taaza Homogenised Toned Milk 1L"
- Store B: "Amul Toned Milk 1000ml pouch"
- Store C: "Amul Taaza Fresh Milk - 1 Litre"
Standard Levenshtein distance fails because volume formatting (1L vs 1000ml) and promotional noise (Super Saver, Combo) distort similarity scores.
Solution: Hybrid Deterministic + LLM Matching
-
Deterministic Normalization:
- Standardizes unit tokens (
ml->l,gm->kg). - Strips common promotional stop-words.
- Extracts brand tokens and item weight/quantity.
- Standardizes unit tokens (
-
Sub-second LLM Fallback (via Groq):
- When deterministic confidence falls below threshold, an ultra-fast LLM call through
@ai-sdk/groqevaluates the ambiguous title candidates in ~200–300ms to verify item equivalence.
- When deterministic confidence falls below threshold, an ultra-fast LLM call through
2. Natural Language AI Mode (Recipe to Cart)
Instead of searching item-by-item, users can type natural meal prompts:
"Ingredients for Butter Chicken and Jeera Rice for 4 people"
Using structured outputs with Groq:
- The prompt translates into a list of normalized grocery ingredients with realistic quantities.
- The auto-matcher searches available inventory across stores and maps them directly to active stock.
3. Architecture & Tech Stack
- Framework: React Native with Expo SDK 58 & Expo Router (typed routing)
- Language: TypeScript
-
AI Integration: Vercel AI SDK (
@ai-sdk/groq) - Animation: React Native Reanimated (fluid price delta transitions)
- Store Bridges: In-app WebViews to preserve session context and export finalized baskets
-
Testing: 560+ deterministic unit tests using
jest-expo(mocked network & store layers)
4. Key Takeaways
- Avoid purely algorithmic matching for messy real-world retail data: Combining lightweight token extraction with an ultra-fast LLM fallback yields over 90%+ match accuracy without breaking latency budgets.
- Mocking third-party integrations early: Quick-commerce web structures change rapidly. Keeping store adapters strictly decoupled and mocked in unit tests saved countless hours of debugging.
Links & Resources
- Google Play Store: Download BasketBuddy
- Video Walkthrough: Watch on YouTube
Have you tackled entity matching across messy e-commerce catalogs? What heuristics or embedding approaches worked best for you? Let's discuss in the comments!
Top comments (0)