DawaiSaathi

---

💛 The Cause

Elderly patients commonly take 4–8 daily medications. Roughly 50% of patients with chronic diseases do not take medicines as prescribed (WHO estimate). Confusion about which pill, when, and with/without food causes hospitalizations that are largely preventable. Furthermore, dangerous interactions go unnoticed across multiple prescriptions, and people overpay for branded molecules because they don't know about generic equivalents like India's Jan Aushadhi program.

Every existing solution assumes a smartphone-literate patient. Pill-reminder apps require the patient to read English, install an app, and respond to push notifications. The highest-risk users (rural, elderly, low-literacy) are excluded.

DawaiSaathi ("Medicine Companion") splits the roles. The caregiver (e.g., an adult child) snaps photos of the medicine strips to set up the schedule. The patient receives an automated phone call in their own language at dose time and simply presses 1 to confirm. No app, no reading required.

✨ What it does

• 📸 Scan & Extract: Photograph up to 5 medicine strips at once. The AI extracts brand name, salt composition, form, MRP, expiry, and manufacturer in seconds.
• 🌐 Caregiver UI languages: Reviewed interface dictionaries are available in English, Hindi, and Spanish. The app only exposes a UI language after the full dictionary is checked in and tested.
• 🗣️ Spoken Reminders (IVR): Places outbound voice calls using Twilio. Hindi and English remain first-class; Bengali, Arabic, French, Portuguese, Afrikaans, Amharic, Swahili, Hausa, Yoruba, and Spanish are available for call setup. Languages without a matching Twilio <Say> locale use generated audio rather than an unrelated fallback voice.
• 🚨 Safety Checks: Cross-references medications against openFDA label data to detect and clearly explain dangerous drug-drug interactions (e.g., Warfarin + Aspirin) without hallucinating.
• 💰 Generic Savings: Identifies identical generic medicines from India's Jan Aushadhi program and shows exactly how much money the patient could save each month.
• 📊 Caregiver Dashboard: A beautiful, accessible Next.js web app that tracks adherence, upcoming doses, and alerts the caregiver if a dose is missed.

---

🏗️ Deployment and data architecture

The clean public origin is dawaisaathi.pages.dev. A lightweight Cloudflare Pages gateway forwards requests privately to the OpenNext Worker, so the stable Pages URL and Worker-only bindings can coexist.

The live deployment currently uses the isolated D1/R2 preview runtime. The Supabase Postgres + Auth + RLS migration is present in this repository but deliberately not claimed as live until a Supabase project has been linked, the migration applied, and tenant/duplicate-delivery tests have passed. This is a safety gate for health information, not a cosmetic switch.

graph TD
    subgraph Data Ingestion
        C[Caregiver browser / Android TWA] --> P[Cloudflare Pages gateway<br/>dawaisaathi.pages.dev]
        P --> A[Next.js App Worker<br/>OpenNext on Cloudflare Workers]
        T[Twilio webhooks] --> A
    end

    subgraph AI & Verification
        A --> B[Groq API<br/>Llama-4-Scout Vision]
        A --> HF[Hugging Face Inference API<br/>VITS / MMS SOTA Voice Engine]
        A --> C2[Gemini / OpenAI TTS<br/>Voice Generation Fallbacks]
        A --> D[openFDA API<br/>Label Grounding]
    end

    subgraph State & Storage
        A --> E[(D1 — current isolated runtime)]
        A --> F[(Private Cloudflare R2<br/>photos + generated audio)]
        A --> I[(Dedicated R2 + Durable Object<br/>OpenNext incremental cache)]
    end

    subgraph Production data cutover
        A -.- S[(Supabase Postgres + Auth)]
        S -.- H[Households / members / medication records]
    end

Read the production architecture and rollout gates before inviting real households.

🚀 Getting Started

The easiest way to try DawaiSaathi is the live dashboard at dawaisaathi.pages.dev.

If you want to run it locally or deploy it yourself:

Prerequisites

• Node.js 22+
• Groq API Key: For fast Llama-4-Scout Vision extraction and LLM routing.
• OpenAI API Key: For GPT-4o-mini-tts generation.
• Twilio Account: For live IVR phone calls (simulated calls work in the browser without Twilio).
• Supabase Project: Required only for the gated Postgres/Auth cutover described in supabase/README.md.

Running Locally

git clone https://github.com/vksh1cool/DawaiSaathi.git
cd DawaiSaathi
npm install

# Setup your .env file
cp .env.example .env
# Edit .env with your Groq, OpenAI, and (optionally) Twilio keys

# Initialise the local D1 database used by `next dev` through OpenNext
npm run d1:migrate:local

# Start the dev server
npm run dev

# In a separate terminal, run the local background worker for reminders
npm run worker

Head over to http://localhost:3000 to access the dashboard.

With DEMO_MODE=true, you can optionally seed the local D1 demo after the app is running:

npm run demo:seed

npm run db:push is still useful for standalone Prisma/SQLite scripts, but it does not initialise the D1 binding that OpenNext uses during next dev.

Deploying to Cloudflare

The production setup has two deployables on purpose: the OpenNext Worker holds the application and private bindings; the Pages project is a minimal gateway that owns the clean public URL.

1. Create the Worker bindings in wrangler.jsonc: D1 (isolated runtime), private R2 media/cache buckets, and the OpenNext cache Durable Object.
2. Set secrets with wrangler secret put (AI keys, Twilio, access-gate secrets, and later Supabase credentials). Never commit them or add them to vars.
3. Deploy the application Worker: npm run cf:deploy.
4. Create a Pages project named dawaisaathi, then deploy pages-gateway: npx wrangler pages deploy pages-gateway --project-name dawaisaathi.
5. Onboard launchpixel.in in Cloudflare Email Sending (SPF/DKIM) and authorize feedback@launchpixel.in. The in-app feedback button sends appreciation or feature requests to contact@launchpixel.in; until this binding is live it fails closed rather than pretending an email was sent.
6. Verify https://dawaisaathi.pages.dev/api/app-info and a real, consented Twilio test call. The live build reports whether telephony is configured.

For the Supabase cutover, follow supabase/README.md first. Do not change the runtime database flag merely because credentials exist.

Twilio trial-mode rule

For this proof of concept, one Twilio sender number may serve every household, while each household's consent and reminder records remain isolated. A Twilio trial account can only contact verified destination numbers: test only with explicit consent and verified recipients. Keep caregiver sign-in OTP delivery separate from medication-reminder traffic, send no medicine names or doses by SMS, and honor STOP across the shared sender before any future message is queued.

---

💬 Product feedback

The floating feedback button asks for one short thing to improve or one short thing the caregiver appreciated and why. It warns people not to include medicine names, prescriptions, phone numbers, or other health details. Messages are sent server-to-server to contact@launchpixel.in; no browser mailto link or email credential is exposed.

---

🛠️ Tech Stack

• Framework: Next.js 15 (App Router) running on Cloudflare Workers (via OpenNext).
• Language: TypeScript (strict).
• Current isolated data runtime: SQLite locally / Cloudflare D1 in the deployed demo.
• Production data target: Supabase Postgres + Auth + RLS, with a migration and rollout gates in supabase/.
• Storage: Private Cloudflare R2 for photos and generated TTS audio; private routes use no-store rather than public object URLs.
• Styling: Tailwind CSS 4.x with a custom accessibility-first design system.
• Framework: Next.js 15 (App Router) running on Cloudflare Workers (via OpenNext).
• Language: TypeScript (strict).
• Data runtime: SQLite locally / Supabase Postgres + Auth + RLS / Cloudflare D1.
• Storage: Private Cloudflare R2 for photos and generated TTS audio; private routes use no-store rather than public object URLs.
• Styling: Tailwind CSS 4.x with a custom accessibility-first design system.
• AI & Voice Engines: Groq API (llama-4-scout for structured extraction, llama-3.3-70b for logic), Hugging Face Inference API (espnet/kan-bayashi_ljspeech_vits & facebook/mms-tts-eng SOTA voice models), Gemini Native TTS & OpenAI TTS fallbacks.
• Telephony: Twilio Programmable Voice (TwiML).
• Android: Trusted Web Activity generated with Bubblewrap; automated signed APK releases (arm64-v8a & armeabi-v7a).

---

📱 Android APK and GitHub Releases

The Android app is versioned from android/twa-manifest.json and opens the production origin https://dawaisaathi.pages.dev. The release workflow produces signed APKs (arm64-v8a & armeabi-v7a) and an Android App Bundle (AAB), attaches SHA-256 checksums, and creates a GitHub Release for each vX.Y.Z tag (current release: v1.1.0).

---

🎯 The Demo

DawaiSaathi includes a built-in demo persona: Kamla Devi. To experience the app as a caregiver setting up medicines for an elderly patient:

1. The UI will guide you through scanning a demo medicine strip, checking interactions, and initiating a simulated phone call.
2. You can hear exactly what Kamla Devi hears in Hindi with warm humanizer voice instructions and confirm the dose by pressing 1.

---

🗺️ Roadmap & Status

Feature	Status	Details
Android APK (TWA)	✅ Live (v1.1.0)	Automated signed release pipeline building arm64-v8a & armeabi-v7a APKs on GitHub Releases.
Supabase Tenant Architecture	✅ Live	Supabase Auth, household RLS security policies, caregiver invitations, and Postgres data layer fully integrated.
Hugging Face SOTA Voice Engine	✅ Live	Free open-source TTS voice models (espnet/kan-bayashi_ljspeech_vits & facebook/mms-tts-eng) with automatic cold-boot retry.
Jan Aushadhi Kendra Locator	✅ Live	Direct integrated store locator linking to https://janaushadhi.gov.in.
WhatsApp Bot Integration	⏳ Planned	Send text-based alerts to the caregiver if a dose is missed.
Pharmacy Ordering	⏳ Planned	Direct integration to re-order medicines when a strip is running low.

Have an idea to improve DawaiSaathi? Open an issue or submit a pull request!