A WhatsApp-based AI health assistant for senior citizens, powered by ASI:ONE. No app to install. No new interface to learn. Just WhatsApp.

🎥 Demo Video: Link

NanaGPT helps elderly individuals manage their daily healthcare through simple WhatsApp messages. It handles medicine reminders with automatic follow-ups, prescription explanations, health question answering, and health log tracking — all in the user's preferred Indian language.

WhatsApp User
     |
     v
Meta WABA Webhook  (POST /webhook)
     |
     v
Flask App  (Render)
     |
     ├── State Machine  (onboarding / menu / reminder / prescription / ...)
     ├── ASI:ONE API    (language understanding, vision, multilingual replies)
     ├── PostgreSQL (Supabase) (user profiles, reminders, health logs, chat history)
     └── APScheduler    (reminder delivery + follow-up nudges, checked every minute)

UptimeRobot pings the deployed URL every 5 minutes to keep the Render free
instance awake, since the scheduler only runs while the instance is alive.

• AI — ASI:ONE (asi1-mini)
• Backend — Python, Flask
• Database — PostgreSQL, hosted on Supabase (free tier)
• Scheduler — APScheduler
• WhatsApp — Meta WhatsApp Business API (WABA)
• Hosting — Render (free tier)
• Uptime Monitoring — UptimeRobot (free tier)

ASI:ONE is the core intelligence layer. It is accessed via an OpenAI-compatible API, integrated using the standard Python openai SDK with a custom base_url.

• Personalized responses — Every API call includes a dynamically built system prompt with the user's name, known conditions, and current medicines. Responses are contextual, not generic.

• Multilingual output — The system prompt instructs ASI:ONE to respond in the user's selected language and correct script (Devanagari for Hindi/Marathi, Gujarati script, Tamil script), with explicit, repeated enforcement to prevent fallback to English.

• Structured data extraction — ASI:ONE embeds structured JSON inside custom tags in its response:

<REMINDER>{"medicine":"Metformin","time":"08:00","frequency":"daily"}</REMINDER>
<HEALTHLOG>{"type":"BP","value":"130/85"}</HEALTHLOG>

These are parsed by the app and written to the database. The tags are stripped before the reply reaches the user.

• Prescription image analysis — Prescription photos are downloaded from WhatsApp, converted to base64, and sent to ASI:ONE as vision input via a data URI (not a direct URL, since WhatsApp media links expire within minutes). The model returns a structured explanation in the user's language.

• Conversation context — The last three exchanges are retrieved from the database and included in every API call, maintaining session context across stateless webhook requests.

nanagpt/
├── app.py                  # Flask app, webhook handler, state machine
├── asi_handler.py          # ASI:ONE API calls, prompt building, tag extraction
├── whatsapp.py             # WABA message sending, media download
├── firebase_handler.py     # PostgreSQL operations (users, reminders, logs, history)
├── reminder_scheduler.py   # APScheduler — due reminders + follow-up nudges
├── requirements.txt
├── render.yaml             # Render deployment config
└── .env                    # Environment variables (not committed)

Note: firebase_handler.py is named for historical reasons from an earlier prototype that used Firebase. It currently contains PostgreSQL (via psycopg) operations.

git clone https://github.com/yourusername/nanagpt.git
cd nanagpt

pip install -r requirements.txt

1. Create a free project at supabase.com
2. Go to Project Settings → Database → Connection String → URI
3. Copy the connection string (use a password with no special characters such as @ or /, since they break URL parsing)

Create a .env file:

ASI_API_KEY=your_asi_one_api_key
ASI_BASE_URL=https://api.asi1.ai/v1
ASI_MODEL=asi1-mini

WHATSAPP_TOKEN=your_meta_permanent_access_token
PHONE_NUMBER_ID=your_whatsapp_phone_number_id
WEBHOOK_VERIFY_TOKEN=your_chosen_verify_token

CAREGIVER_PHONE=91XXXXXXXXXX

DATABASE_URL=postgresql://postgres:[password]@db.xxxxxxxx.supabase.co:5432/postgres

python app.py

ngrok http 5000

Use the generated https:// URL as your webhook in the Meta Developer Portal.

1. Push the repository to GitHub
2. Go to render.com → New Web Service → Connect repository
3. Set build command: pip install -r requirements.txt
4. Set start command: gunicorn app:app --bind 0.0.0.0:$PORT --workers 1
5. Add all environment variables from .env, including DATABASE_URL, in the Render dashboard
6. Deploy — your webhook URL will be https://your-app.onrender.com/webhook

Render's free tier spins the instance down after 15 minutes of inactivity, which stops the background scheduler entirely.

1. Go to uptimerobot.com → create a free account
2. New Monitor → HTTP(S) → URL: https://your-app.onrender.com/
3. Interval: every 5 minutes

This keeps the instance continuously awake so reminders and follow-up nudges fire at the correct time.

1. Create an app at developers.facebook.com → Business type
2. Add the WhatsApp product
3. Under WhatsApp → Configuration, set:
   • Callback URL: https://your-app.onrender.com/webhook
   • Verify Token: value of WEBHOOK_VERIFY_TOKEN in your .env
4. Subscribe to the messages webhook field
5. For a permanent access token (the default token expires every 24 hours), create a System User in Meta Business Manager with whatsapp_business_messaging permission

First message (Hi)
  └── Onboarding
        ├── Name
        ├── Age
        ├── Known conditions
        ├── Current medicines
        └── Prescription photo (optional)
              └── Main Menu

Main Menu
  ├── 1 - Set Medicine Reminder
  ├── 2 - View My Reminders
  ├── 3 - Read Prescription
  ├── 4 - Ask Health Question
  ├── 5 - Log Health Reading
  ├── 6 - View Health History
  └── 7 - Change Language

Reminder Flow
  ├── Initial reminder sent at scheduled time
  ├── +10 min, if no "done" → Follow-up 1
  ├── +10 min, if no "done" → Follow-up 2 (final)
  └── User reply "done" at any point → clears all pending follow-ups

Reply 0 anytime → Main Menu

flask==3.0.3
openai==1.30.5
httpx==0.27.0
requests==2.32.3
APScheduler==3.10.4
python-dotenv==1.0.1
gunicorn==22.0.0
Pillow==10.4.0
psycopg[binary]==3.3.4
pytz==2024.1

A non-exhaustive log of real issues hit while building and deploying, kept here for anyone extending this project:

• 401 Authentication Error on every outbound message — Meta's temporary access token expires every 24 hours. Fixed by generating a permanent token via a System User in Meta Business Manager.
• TypeError: Client.__init__() got an unexpected keyword argument 'proxies' — version mismatch between openai and httpx. Fixed by pinning openai==1.30.5 and httpx==0.27.0.
• undefined symbol: _PyInterpreterState_Get — psycopg2-binary has no compiled wheel for Python 3.14. Fixed by switching to psycopg[binary] (psycopg3).
• ImportError: cannot import name 'db' — leftover Firebase-style import after migrating to SQL, caught by keeping all DB access behind the same function names (get_user, upsert_user, etc.) regardless of the underlying engine.
• Reminders saved but never delivered — root cause was twofold:
  1. APScheduler was started inside if __name__ == "__main__":, which Gunicorn never executes. Fixed by starting the scheduler at module load.
  2. Render's free tier sleeps after 15 minutes idle, stopping the scheduler entirely. Fixed with an UptimeRobot monitor pinging every 5 minutes.
• Profile reset and reminders disappearing every session — Render's free-tier filesystem is ephemeral; the original SQLite file was wiped on every redeploy/restart. Fixed by migrating to a hosted PostgreSQL database on Supabase.
• failed to resolve host on DB connection — password containing @ broke the connection string parsing. Fixed by resetting the Supabase DB password to one without special characters.
• Prescription reading silently failing — WhatsApp media URLs expire within minutes and require an auth header to fetch, which ASI:ONE's vision endpoint can't supply. Fixed by downloading the image server-side and sending it to ASI:ONE as a base64 data URI instead of a raw link.
• Numeric menu replies misrouted (e.g., selecting language option "2" was interpreted as main menu option "2") — fixed by making state-based routing strictly take priority over any number/keyword detection.
• TypeError: 'datetime.datetime' object is not subscriptable — PostgreSQL returns timestamp columns as native datetime objects, not strings (unlike SQLite). Fixed by formatting with .strftime() instead of string slicing.

• The Meta temporary access token expires every 24 hours. Use a System User permanent token for any deployment beyond local testing.
• ASI:ONE's vision capability is used for prescription reading. Ensure images sent by users are reasonably well-lit and in focus for accurate results.
• On WhatsApp's free/test tier, outbound messages can only be sent to a user within a 24-hour window after they last messaged the bot. This window resets every time the user sends a new message.
• Render's free tier is sufficient for demos and testing but is not suited for reliable production-grade scheduled delivery without an external uptime monitor keeping the instance awake.

• Tejas Gadge
• Anisha Shankar