The problem
Laptop shopping is a classic "too many filters, not enough judgment" problem — buyers don't think in spec sheets, they think in "I need something fast for video editing under $1500." Filter-based product search can't answer that. The brief was a recommendation engine that understands intent, not just keywords, and responds fast enough to feel conversational.
My role
I built the full retrieval and recommendation pipeline — data ingestion, embedding generation, hybrid ranking, and the conversational layer on top — as a solo engineer.
What I built
A RAG-based recommendation system that takes a natural-language need and returns personalized laptop recommendations in under 200ms, by combining three different signal sources instead of relying on one:
- Benchmark data — actual performance numbers, not marketing specs.
- Editorial content — review and comparison context that a spec sheet can't capture.
- Semantic search — embeddings over product data so "good for video editing" maps to the right hardware characteristics even when those exact words never appear in the listing.
Underneath that:
- Webhook-driven ingestion keeps benchmark and product data current without manual re-syncing.
- Hybrid ranking blends raw performance metrics with editorial signal so the engine doesn't just recommend the fastest chip, it recommends the right machine for the stated use case.
- Sub-200ms responses via Sentence Transformers embeddings + pgvector similarity search and a Redis caching layer in front of the hot paths.
Tech stack
FastAPI for the API layer, PostgreSQL + pgvector for vector storage, Redis for caching, GPT-4o and Gemini Flash for the conversational layer, Sentence Transformers for embeddings, Sanity CMS for editorial content, and Docker + AWS S3 for infra and assets.
Why this matters if you're hiring
If your product needs search or recommendations that understand what users mean, not just what they typed, this is direct evidence I can build that — including the parts people usually skip, like keeping retrieval fast and keeping the underlying data fresh without manual babysitting.