SE7A

AI food & fitness coach — marketing site, waitlist, web product, and a native iOS/Android app via Expo. All three vision features are live (plate / menu / body) on both surfaces.

Built on Next.js App Router. Deployed to Vercel. Data + auth in Supabase. Vision via Vercel AI Gateway → Anthropic Claude.

Repo layout

/                          ← Next.js web app (marketing + product)
apps/
  mobile/                  ← Expo native app (see apps/mobile/README.md)
supabase/migrations/       ← SQL — shared between web and mobile

The web's /api/* routes accept either session cookies (web caller) or Authorization: Bearer <access_token> (mobile caller), so the two surfaces share the same backend.

Routes

/                       ← marketing landing + waitlist
/privacy                ← privacy policy
/login                  ← magic-link sign in
/onboarding             ← first-login profile form (gated)
/dashboard              ← targets + today's ledger + weight log (gated)
/scan/plate             ← camera upload + macro review + add-to-log (gated)
/scan/menu              ← menu photo + dish ranking against today's budget (gated)
/scan/body              ← body-fat range + weeks-to-goal; photo NOT stored (gated)
/auth/callback          ← Supabase magic-link return
/api/profile            ← POST onboarding, computes + saves targets
/api/weight             ← POST weight log, retunes targets
/api/scan/plate         ← POST plate image, runs vision, persists scan
/api/scan/menu          ← POST menu image, runs vision + budget-aware ranking
/api/scan/body          ← POST body image, runs vision; image NEVER stored
/api/ledger/add         ← POST selected items to today's log
/api/ledger/today       ← GET today's totals + remaining vs targets

Route groups: app/(marketing) holds the public site, app/(app) holds the gated product surface. Both render at the root URL — route groups don't add URL segments.

Local dev

npm install
cp .env.example .env.local       # fill in Supabase + AI Gateway keys
npm run dev

Requires Node ≥ 20.6.

Run the migrations once against your Supabase project (paste into SQL editor):

supabase/migrations/0001_phase0.sql   ← profiles, weight_logs, RLS, triggers
supabase/migrations/0002_phase1.sql   ← scans, meal_items, plate-scans bucket
supabase/migrations/0003_phase2.sql   ← menu-scans bucket

Enable magic-link auth in Supabase: Auth → Providers → Email → Magic Link. Add redirect URLs (Auth → URL Configuration → Redirect URLs) for both http://localhost:3000/auth/callback and https://se7a.app/auth/callback.

Get an AI Gateway key from vercel.com/dashboard → AI Gateway → API Keys and put it in .env.local as AI_GATEWAY_API_KEY. On Vercel deployments this is automatic via OIDC — no key needed in env vars.

Tests

npm test            # node --test, tsx loader — macro math coverage
npm run typecheck   # tsc --noEmit
npm run build       # production compile

lib/macros.ts is pure functions; lib/macros.test.ts covers BMR (Mifflin-St Jeor), TDEE multipliers, kcal-floor clamps, lean-mass refinement, and the full computeTargets integration. 18 tests, all green.

Deploy

Push to main → Vercel auto-deploys.

Required env vars on Vercel (Production + Preview):

• NEXT_PUBLIC_SUPABASE_URL
• NEXT_PUBLIC_SUPABASE_ANON_KEY

AI_GATEWAY_API_KEY is auto-injected on Vercel; no manual config needed.

Project shape

app/
  (marketing)/                public landing + privacy
  (app)/                      gated: /login, /onboarding, /dashboard, /scan/plate
  auth/callback/              magic-link return
  api/
    profile/                  POST onboarding
    weight/                   POST weight log + retune
    scan/plate/               POST plate image → AI → persist
    ledger/{add,today}/       POST log items, GET today's totals
components/                   shared (Waitlist)
lib/
  supabase/                   SSR + browser + middleware clients
  schemas/                    Zod input + scan-result schemas
  prompts/                    versioned prompt strings (plate.v1, …)
  macros.ts                   BMR, TDEE, macro split, age math
  macros.test.ts              node:test coverage
  ai.ts                       AI Gateway model registry + prompt versions
  ledger.ts                   today's totals + remaining-budget helpers
middleware.ts                 refreshes Supabase session each request
supabase/migrations/          SQL — run against the prod Supabase project

Brand rule, baked into every scan

Outputs are ranges, never point values. Every macro returned from a vision scan has a _low and _high. The model is prompted to surface "invisible costs" (oil, butter, hidden sugars). Confidence is enforced as low | medium | high. A photo can't see the oil; we don't pretend it can.

Roadmap

• ✅ Phase 0 — auth, profile, macro math, dashboard skeleton.
• ✅ Phase 1 — plate scan: image upload → AI Gateway → ledger entry.
• ✅ Phase 2 — menu scan: OCR + budget-aware ranking (order/consider/skip).
• ✅ Phase 3 — body composition scan: honest BF range + weeks-to-goal. Photo analyzed in memory and discarded — never stored (privacy policy commitment).
• ✅ Phase 4 — Expo native app (apps/mobile). Same backend, native camera, magic-link via se7a:// deep links. EAS-ready for TestFlight; see apps/mobile/README.md.