Health is the worst place for an LLM to guess. Every claim must trace to a retrieved source; if it can't, the app declines.

• 45% fewer ungrounded answers (1,120 → 615 failures) across 10,000+ test cases
• p95 latency ~1,950 ms → under 800 ms (Locust, 50,000 requests, 100 concurrent users)
• 60% fewer redundant API calls (200 → 80) over 30 days of logs

Next.js Node.js Grok API vector search IndexedDB Firebase
Status: stabilizing for a public release, with known issues tracked in the repo · Architecture & case study
_{Informational tool, not medical advice.}

Scan text and speak in one flow, with recognition running on the device instead of a cloud round-trip.

• 60% faster (median ~2.0 s → ~0.8 s vs. a cloud OCR baseline), 500+ runs per device on Pixel 5 and OnePlus Nord
• Mode-switch drop-off 32% → 13% (−19 points), Firebase Analytics funnel

TypeScript Next.js Google ML Kit Firebase Cloud Messaging
Status: stabilizing for a public release, with known issues tracked in the repo · Architecture & case study

• Measure first. Every number has a stated method and sample size. Small samples and non-randomized comparisons are labeled as such in the repos.
• Document failures. MyHealthApp answered 20 of 100 questions even when retrieval found nothing relevant. The test suite caught it, I added a hard grounding rule, and it dropped to 7 of 100.
• Prefer refusing to guessing when the cost of a wrong answer is high.

Languages: TypeScript · JavaScript
Frontend: React · Next.js · Tailwind CSS
Backend & data: Node.js · Firebase · IndexedDB (offline-first sync)
AI: LLM APIs · retrieval-augmented generation · on-device ML (Google ML Kit) · evaluation test suites
Quality: Vitest · Locust load testing · GitHub Actions