Avighna Clinical AI
A clinical symptom checker I took from first prototype to 50,000+ users in 18 months. It asks about symptoms, ranks the likely conditions and points people to the right next step, inside apps they already use. In real-world use at three medical centres it reached 72% diagnostic accuracy.
- My role
- Product lead, 0→1: from the first prototype to distribution and commercial plan
- Team
- Started solo, grew to three
- Distribution
- MyJio (400M+ users), JioHealthHub, KiviHealth, ABDM and WhatsApp
- Languages
- Hindi, English and regional languages
The Problem
India has too few doctors for its patients, most of all in Tier 2 and 3 cities, and people often don't know whether a symptom needs a doctor. The aim was a preliminary assessment that was accurate enough to trust, safe enough to put in front of consumers, and available where people already were.
What I Built
- A hybrid engine. A clinical decision-support system combined with retrieval-augmented generation, fine-tuned on 3,000+ verified medical consultations.
- Safety built in from day one. Bias testing, hallucination detection and guardrails, and every result ends with advice to consult a physician.
- A scope people can trust. Symptom checking and triage: likely conditions and a next step, with the diagnosis left to a doctor.
- Distribution, not a new app. Integrations with MyJio, JioHealthHub, KiviHealth and the government's Ayushman Bharat Digital Mission, plus a WhatsApp bot for people who prefer chat.
MyJio / JioHealthHub Integration Flow
End-to-end user flow from MyJio app launch to triage result delivery.
MyJio App
User opens MyJio (400M+ users)
Health Mini App
Navigates to Health section, sees Avighna banner
JHH Authentication
PIN verification via JioHealthHub
Profile Selection
Selects self or family member profile
Symptom Input
Enters up to 5 symptoms via search
Triage Result
Ranked conditions with actions: New DDA, Consult Doctor, Download Report
WhatsApp Healthcare Bot
The same assessment as a WhatsApp bot, for people who prefer chat to an app: Hindi, English and regional languages, available around the clock.
LLM + Clinical Decision Support
Led benchmarking of Avighna's Clinical Decision Support System combined with Large Language Models for enhanced disease prediction. Managed data preparation across 3 model variations, evaluation, and performance reporting. The hybrid RAG approach showed improved diagnostic accuracy compared to standalone approaches.
Decisions and Trade-offs
- Distribution over acquisition. Building inside MyJio and JioHealthHub instead of a standalone app put Avighna in front of 400M+ existing users, at the cost of working within each partner's flow and authentication.
- Triage, not diagnosis. Limiting the product to likely conditions and a next step kept it clinically useful while staying on the right side of regulation and patient safety.
- Safety before scale. Hallucination detection and bias testing were in the first version, not added after growth.
- Fix the first session. Systematic A/B tests of the UX flows, balanced against the safety prompts, cut first-session bounce from 45% to 20%.
Results
I also wrote the five-year commercial plan, including market sizing and the unit economics of Diagnosis-as-a-Service.