GlucoVision — Offline-first explainable diabetic retinopathy screening for rural PHCs

Next.js 16 frontend + Cloudflare Workers API + D1/R2 + NVIDIA NIM fallback
Screen, explain, refer, and follow up from one shared clinical workflow.

Prevent blindness before it starts.
Offline-first • Explainable • eSanjeevani-ready • PHC Edition

🌐 Live Demo — glucovision.pages.dev • API Health • Quickstart • Architecture

Smart India Hackathon 2026 — PHC Edition. A clinical screening suite for diabetic retinopathy that works with an ophthalmoscope + phone, keeps PHCs offline-first, and makes every result explainable with Grad-CAM so ASHA workers, Medical Officers, and ophthalmologists can verify what the model saw.

Keywords for SEO: diabetic retinopathy, AI screening, fundus photography, rural health, PHC, ASHA, offline-first, explainable AI, Grad-CAM, eSanjeevani, telepharmacy, Next.js, Cloudflare D1, NVIDIA NIM, Smart India Hackathon

---

✨ Why GlucoVision

Rural India has	GlucoVision gives
100M+ diabetics, 70% rural never screened	One phone + ophthalmoscope — clip-on optional, no fundus camera procurement
No internet in villages	Offline-first — on-device CNN <2.1s, encrypted local storage, SMS queued, syncs later
No ophthalmologist at PHC	Explainable AI — 5-stage CNN (APTOS/IDRiD) + Grad-CAM heatmap + confidence + plain language → eSanjeevani referral in one tap
Lost to follow-up	Continuity — glucose trends + eye + foot + telepharmacy in one record, PHC dashboard with coverage & camp planning

---

🎥 Demo

• Landing: Cinematic hero, gapless bento, GSAP pinned scroll, horizontal accordions, testimonial carousel — http://localhost:3000/
• App: Login → Dashboard → Patients → Screening (live camera + blur gate + NIM) → Referrals → Telepharmacy → Foot
• Demo logins (demo123): asha@glucovision.in (ASHA), mo@glucovision.in (MO), eye@glucovision.in (ophthalmologist), pharma@glucovision.in, admin@glucovision.in — works offline without backend

---

🏗️ Architecture — UI ↔ API with offline fallback

flowchart LR
  subgraph users[PHC users]
    A["ASHA workers\nMedical Officer\nOphthalmologist\nPharmacist"]
    B["Phone browser\nPWA camera"]
  end

  subgraph frontend[Next.js 16 frontend]
    C["src/app routes\nlogin, signup, dashboard, patients\nscreening, referrals, pharmacy, foot"]
    D["src/components + src/lib/store.tsx\noffline mock + localStorage + API hydration"]
  end

  subgraph worker[Cloudflare Worker API]
    E["Hono routes in backend/src/index.ts"]
    F["Auth\nsignup, login, me, logout"]
    G["Patients, referrals, pharmacy, visits"]
    H["Case chat\nprompt window management"]
    I["NIM inference\n3-model fallback"]
  end

  subgraph cloudflare[Cloudflare data services]
    J["D1 SQLite\npatients, glucose, visits, referrals\npharmacy, auth, chats, secrets"]
    K["R2 fundus images"]
  end

  L["NVIDIA NIM"]

  A --> B --> C
  C -->|HTTPS| E
  D --> C
  E --> F --> J
  E --> G --> J
  E --> H --> J
  E --> K
  I --> L
  E --> I

Loading

Frontend (src/) — Next.js 16 App Router, offline mock/store fallback, live camera screening with blur gating, explainable AI review, patient management, referrals, telepharmacy, foot screening, and case chat. The current route set is login, signup, dashboard, patients, screening, referrals, pharmacy, and foot.

Backend (backend/) — Hono Worker with auth, patient records, referral and pharmacy workflows, case chat, encrypted secrets, D1 persistence, optional R2 image storage, and NVIDIA NIM inference with fallback across multiple models.

Cloudflare DB — D1 (SQLite): backend/migrations/0001_initial.sql through 0007_foot.sql, plus backend/migrations/0004_auth.sql for users and PBKDF2 demo hashes. backend/seed.sql carries the current demo dataset for patients, referrals, and pharmacy records.

---

🚀 Quickstart

1) Frontend (offline mock — no backend needed)

git clone https://github.com/manish-9245/GlucoVision.git
cd GlucoVision
npm install
npm run dev
# http://localhost:3000
# Login: asha@glucovision.in / demo123

2) Full stack with Cloudflare D1 + R2 + NIM (local)

cd backend
npm install

# D1 + R2
npx wrangler d1 create glucovision
# copy database_id into backend/wrangler.jsonc d1_databases[0].database_id
npx wrangler r2 bucket create glucovision-images # or skip — upload falls back to data URL

# Secrets (never commit plain .env)
echo "RANDOM_32B_BASE64" | npx wrangler secret put ENCRYPTION_KEY
echo "RANDOM_32B_BASE64" | npx wrangler secret put JWT_SECRET
echo "nvapi-..." | npx wrangler secret put NVIDIA_API_KEY

# Migrate + seed
npx wrangler d1 migrations apply glucovision --local
npx wrangler d1 execute glucovision --local --file=./seed.sql
# (fix seed.sql: remove PRAGMA/BEGIN if `SQL BEGIN TRANSACTION` error — already fixed)

npx wrangler dev --persist-to=./.wrangler/state --port 8788
# Worker http://localhost:8788 — test: curl http://localhost:8788/api/health

# In repo root .env.local:
# NEXT_PUBLIC_API_URL=http://localhost:8788
cd ..
npm run dev

3) Deploy

# Backend
cd backend
npx wrangler d1 migrations apply glucovision --remote
npx wrangler d1 execute glucovision --remote --file=./seed.sql
npx wrangler deploy
# → https://glucovision-api.<subdomain>.workers.dev

# Frontend — Cloudflare Pages (OpenNext) or Vercel
cd ..
npm run build
npx opennextjs-cloudflare build
npx opennextjs-cloudflare deploy
# or: wrangler pages deploy .open-next/assets --project-name glucovision
# Set env in Pages: NEXT_PUBLIC_API_URL=https://glucovision-api.<subdomain>.workers.dev

---

🔬 AI — NVIDIA NIM, 3-model fallback, keeps import requests

Worker backend/src/nim.ts:18 mirrors Python backup backend/nim_backup.py:55 — exact payloads you provided:

1. nvidia/nemotron-3-nano-omni-30b-a3b-reasoning stream False reasoning_budget 16384
2. moonshotai/kimi-k3 stream True reasoning_effort max
3. meta/llama-3.2-90b-vision-instruct stream False

# backend/nim_backup.py — keep import requests
import requests
invoke_url = "https://integrate.api.nvidia.com/v1/chat/completions"
headers = {"Authorization": "Bearer $NVIDIA_API_KEY", "Accept": "application/json"}
payload = {"messages": [{"role":"user","content":[{"type":"text","text":prompt},{"type":"image_url","image_url":{"url":image_url}}]}], "model": "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning", ...}
response = requests.post(invoke_url, headers=headers, json=payload, stream=False)

Worker callOne() src/nim.ts:67 does fetch(INVOKE_URL) with 85s abort, collects text/event-stream like for line in response.iter_lines(): print(line.decode("utf-8")) vs response.json(). POST /api/nim/infer loops for (m of MODELS) — if one fails other must work, audit to nim_requests D1, POST /api/nim/infer-stream proxies stream.

Frontend src/app/app/screening/page.tsx:306 runInference — converts preview blob: → FileReader dataURL → fetch ${API}/api/nim/infer {image_url, prompt, patientId} → parses stage/confidence else falls back to simulateInference(patient.riskScore).

Encrypted creds backend/src/crypto.ts:1 AES-GCM 256 Web Crypto, storeEncryptedSecret / getEncryptedSecret D1 secrets table (backend/migrations/0002_secrets.sql), Python encrypt_and_save_all_creds() src/nim_backup.py:32 Fernet → .env.encrypted (base64 fallback). Never commit plain .env (see .gitignore).

---

📸 Camera + Blur — instant flag, smooth, functional everywhere

src/lib/blur.ts:4 estimateBlurScore — downscale to 160px, grayscale luminance, Laplacian [0,1,0;1,-4,1;0,1,0], variance → quality 0-100 (<18 very-blurry 22-48, <55 blurry 48-68, <130 ok 68-86, else sharp). src/app/app/screening/page.tsx:272 live loop requestAnimationFrame 120ms (was 320ms) → setLiveBlur + setQuality instantly, border-[3px] border-red-500 shadow-red when isBlurry vs border-emerald-500 when ≥80, banner Too blurry • 42/100 var 12.3 / Hold steady. Capture disabled if liveBlur?.isBlurry, quality<60 blocks inference. Same on foot src/app/app/foot/page.tsx:128. Upload computeQualityFromFile uses real blur, not random.

---

💬 Discuss case — state-of-art prompt + context management

src/components/CaseChat.tsx:1 ↔ backend/src/chat.ts:7 SYSTEM_PROMPT (8 rules: explainable lesions + heatmap, stage 0-4 strict, confidence + protocol, no auto-prescription, red flags, Hindi switch, JSON {summary,findings,stage,confidence,lesions,gradcam_note,next_step,referral,disclaimer,follow_up} + markdown, blur gate). Few-shot 2 examples (Moderate exudates, No DR). Context patientContextBlock() injects 28 real patients: age/gender/village, diabetesYears/type, HbA1c, BP, risk, symptoms, meds, glucose trend last4, prior visits, lastScreened, foot + visit.image_url. Window MAX_HISTORY_TURNS=8, MAX_PROMPT_CHARS=9000, sliding window drops oldest pair if over budget. UI shows preview thumbnail + Include image toggle, history max-h-[320px], POST /api/cases/:patientId/chat persists case_chats D1.

---

🔐 Auth — end-to-end coherent

backend/migrations/0004_auth.sql:1 users(id,name,email,password_hash,salt,role phc village) PBKDF2 100k src/auth.ts:12, JWT HS256 src/auth.ts:28, POST /api/auth/signup|login, GET /api/auth/me, POST /api/auth/logout, GET /api/auth/users. Seeded 5 demo demo123 with precomputed salts/hashes. Frontend src/lib/auth.tsx:1 AuthProvider — localStorage gv_token/gv_user, USE_API → fetch /api/auth/me else offline mock (role inferred from email prefix), src/components/AppShell.tsx:22 guards if (!loading && !user) router.replace("/login"), shows Logo + user phc/village + LogOut. Flow / → /login → /app/dashboard → Patients → Screening (camera+blur+NIM) → Referrals → Pharmacy → Foot.

---

🖼️ Images — all related, local public/images

No picsum. All public/images local, Google Search via Category:Rural health in India Wikimedia Commons (Bharat Nirman eye/blood camps, Govt of India, public domain) + NIH fundus + Indian portraits. See src/app/page.tsx:331 fundus-proliferative.jpg etc. grep -r "picsum|unsplash|wikimedia" src → 0.

---

📦 Open Source Level

• License LICENSE MIT, Contributing CONTRIBUTING.md, Code of Conduct CODE_OF_CONDUCT.md, Security SECURITY.md
• Issues templates .github/ISSUE_TEMPLATE/*, PR template .github/pull_request_template.md, CI .github/workflows/ci.yml (npm run build + tsc)
• SEO: repo description, topics, public/og-image.png (fundus + concentric retina), public/manifest.json, sitemap via Next, robots.txt, og:* meta in src/app/layout.tsx
• No leaks: .gitignore ignores .env*, .wrangler, .env.encrypted, node_modules; secrets only via wrangler secret put + D1 secrets (AES-GCM), never in repo (verified grep -r "nvapi-" → only nim_backup.py startswith check, no real key)

---

🛠️ Scripts

Script	What
npm run dev	Next dev 3000 (offline mock if no API)
npm run build	Next build (13 routes including /login /signup)
npm run backend:dev	wrangler dev --persist-to Worker :8788
npm run backend:deploy	wrangler deploy
npm run backend:migrate	wrangler d1 migrations apply local+remote
npm run backend:seed	wrangler d1 execute --file=./backend/seed.sql

---

👥 Credits

Built for Smart India Hackathon 2026 — PHC Edition. Logo concentric retina logos/iterations/iteration-2.svg (ink #0a0a0f + teal #0f766e, 2 colors, 512×512). PWA manifest.json + sw.js, typography Cabinet Grotesk (display) + Geist (body).

---

📄 License

MIT — see LICENSE.

---

_{Built with ❤️ for Bharat — offline-first, explainable, eSanjeevani-ready. Star ⭐ if useful!}