FindMeCafe
FindMeCafe turns a plain-language description — "quiet, good wifi, outdoor seating" — into ranked café results on a map, using LLM intent extraction and a caching layer that stretches a single places API call across roughly 30 searches.
| Role | Product build — LLM intent extraction, caching pipeline, map UI |
|---|---|
| Stack | LLM intent extraction · Serverless architecture · Cached places pipeline · Map UI |
| Category | SEMANTIC-UI |
| Status | LIVE |
The problem
Nothing could find you a café from plain language and vibe — the closest option was pasting queries into ChatGPT: a broken experience scattered across tools instead of living under one roof.
The intent layer
An LLM turns the free-text query — "quiet, good wifi, outdoor seating" — into structured preferences (ambience, amenities, distance) instead of matching keywords against a database. That structured output is what actually drives the search, not the raw sentence.
The caching layer
A cached places pipeline stretches a single upstream places-API call across roughly 30 searches, so ranked results land as map pins with AI-written summaries without paying the per-query cost of hitting the places API directly on every search.
Why the caching mattered
Keeping inference and API costs low enough that natural-language search didn't break the unit economics drove most of the architecture. The caching layer exists specifically so semantic search stays serverless and cheap to run at scale, not just fast to build for a demo.
The outcome
A live, working product the founder could hand to real users from the first sprint. In the founder's own words below.
I came in with an idea, and they shipped a real product — a genuinely great, responsive UI with quality you could feel from the first sprint. Deliverables landed on time, every single time.
