AI-Powered Smart Shopping Assistant
Period: Mar 2024 – Jul 2024
Type: Startup concept / Personal project
Problem
Online shopping still relies on keyword search. A user looking for “something to keep my coffee hot at my desk” has to figure out the right product category and keywords themselves. LLMs understand intent, but naively feeding a product catalog into an LLM is too slow and too expensive for real-time recommendations.
What I Built
An intelligent shopping assistant that translates natural language needs into product recommendations without keyword dependency.
What Was Novel
The core challenge was latency. A user query like “I need a gift for a runner who likes tech” could map to multiple product categories (GPS watches, wireless earbuds, fitness trackers). Searching all categories sequentially is too slow.
My solution had two parts:
Parallel category retrieval via intent classification. A small fine-tuned LLM classifies user intent into product categories in a single forward pass. This triggers parallel retrieval across all matching categories simultaneously, rather than sequential search. The result: more products surfaced per query with lower latency.
Pre-computed category embeddings with cached mappings. Category-to-product mappings are pre-computed and cached, so retrieval starts the moment categories are identified rather than waiting for full query embedding. This cut perceived latency significantly compared to a standard RAG pipeline.
Technical Stack
- Fine-tuned small LLM for intent classification
- Custom RAG system with pre-computed embeddings
- Vector similarity search for product matching
- LLM-powered response generation