1// Baymax.md
2// ---------------------------
3
4# Baymax
5
6Private group health companion — medical record vault, biomarker tracking, and an AI that reads your full history to surface what matters.
7
8**TYPE:** Web Application (PWA)
9**STATUS:** Live
10**LIVE:** https://baymaxos.vercel.app
11**DEV:** http://localhost:4000
12**WORKSPACE:** ~/GitHub/baymax
13**GITHUB:** nikanjuri/baymax (private)
14
15---
16
17## FEATURES
18
19- Daily check-ins and longitudinal health tracking across biomarkers and body composition
20- Medical record vault with private storage and row-level security on every user-scoped table
21- AI biomarker extraction — upload a lab report and Gemini structures the values
22- Smart-scale OCR for body-composition readings
23- Longitudinal AI health analysis over your full history via Claude
24- Faithfulness eval harness — Vitest golden-set contracts verify pipeline assembly and output coverage (no live API in CI)
25- Source-attributed findings with provenance back to structured health records
26- AI-assisted meal planning
27- Installable as a PWA
28
29---
30
31## TECHNOLOGIES
32
33- Next.js 16 (App Router, Turbopack)
34- TypeScript
35- Supabase (Postgres, Auth, Storage, RLS)
36- Anthropic Claude Opus 4.7 (longitudinal analysis)
37- Google Gemini 3 Flash / 3.1 Pro (extraction + OCR)
38- Vitest (faithfulness eval harness)
39- shadcn/ui + Tailwind CSS 4
40
41---
42
43## NOTES
44
45Baymax is a health companion for a private group of friends: track check-ins, vault medical records, watch biomarkers and body composition over time, and let an AI that has read your full history surface what's worth your attention.
46
47### Data in, structure out
48
49Lab reports go in as PDFs or photos; Gemini extracts and structures the biomarker values so trends are chartable immediately. Smart-scale photos get the same OCR treatment for body-composition tracking. Everything lands in Supabase with row-level security on every user-scoped table.
50
51### Analysis that knows your history
52
53The analysis layer uses Claude over the full longitudinal record — not a single report in isolation — so it can flag drift across months of data, connect biomarkers to check-in patterns, and suggest what to watch next. Findings are emitted as discrete, source-attributed claims so each conclusion traces back to structured rows in your vault — not free-floating prose.
54
55### Faithfulness you can regression-test
56
57A Vitest eval harness exercises the analysis pipeline with anonymized golden-set fixtures: assembly contracts assert every required input fact reaches the model, and mechanical coverage checks catch dropped or hallucinated references. Runs in ~0.2s with no live API calls — so prompt and model changes get deterministic regression detection before they hit users. A meal-planning assistant builds on the same longitudinal context.
58
59### Built for daily use
60
61It installs as a PWA, keeps the interface calm (shadcn/ui on Tailwind 4), and is hosted on Vercel with Supabase in the Mumbai region for low-latency access.