Clinic Scout AI

Inspiration People often need care quickly and do not have time to call ten places just to find one that is open, walk-in friendly, and reachable. Clinic Scout AI was inspired by that real gap: location-based results exist, but trustworthy, actionable next steps are still hard to get in one place.

What it does It helps users find nearby clinics and recommends the best option for their urgency level by combining:

Nearby clinic discovery Smart ranking (distance, open status, walk-in suitability, reachability) Website fact-checking for key details like hours, walk-ins, and booking/contact options Transparent reasoning with clear Unknown values when data is not verified Real-time progress streaming so users can follow each step How we built it Clinic Scout AI is built as a Next.js app with a tool-using Gemini agent orchestrating the search flow.

Frontend: React-based UI with urgency-aware states, recommendation view, and live progress API layer: server route that streams Server-Sent Events for step-by-step updates Agent loop: Gemini chooses which tools to call, when to inspect more clinics, and when to finalize Tooling pipeline: Geocode user location Search clinics from OpenStreetMap and Overpass Filter specialty-only listings Rank candidates with a deterministic waterfall Inspect clinic websites and verify evidence quotes Finalize recommendation with guardrails Safety architecture: model cannot invent clinic facts; final recommendations are validated against verified fields Reliability: caching, retry logic, stale-cache fallback, and deterministic fallback mode when AI is unavailable Challenges we ran into Free-tier AI quota limits and intermittent model/API availability Overpass and external directory instability under load Preventing hallucinated medical claims while still using an LLM for orchestration Handling messy real-world website hour formats without making unsafe open-now guesses Maintaining useful recommendations even when data is incomplete or partially unknown Accomplishments that we're proud of Strong fact firewall: recommendations can only cite verified fields Graceful degradation: app still works when AI is down or quota is exhausted Transparent UX: users see exactly what is known, unknown, and why Robust behavior under unreliable upstream services via cache and fallback strategy Comprehensive automated testing with full passing coverage, including happy-path and safety-critical scenarios What we learned In healthcare-adjacent products, correctness and traceability matter more than fluent model output Agentic systems need hard guardrails, not just prompt instructions Unknown is often the safest truthful answer when evidence is missing Deterministic ranking plus AI orchestration gives a practical balance of reliability and flexibility Streaming progress significantly improves trust and perceived performance What's next for Clinic Scout AI Expand data quality with deeper but still bounded site inspection Improve relevance classification beyond keyword-only methods Add richer temporal handling for holidays and special-hour exceptions Introduce user feedback loops to refine ranking quality over time Strengthen observability and analytics for production reliability and safety monitoring Explore multilingual support and broader region coverage Built With Next.js 16 React 19 TypeScript 5 Tailwind CSS 4 ESLint 9 Gemini API (tool-calling agent orchestration and extraction) OpenStreetMap and Overpass API (clinic discovery) Nominatim (geocoding) Server-Sent Events for real-time streaming updates