Building MannSaathi — An AI Health Companion That Understands Fear, Not Just Symptoms

#ai #nextjs #machinelearning #healthtech

# Building MannSaathi — An AI Health Companion That Understands Fear, Not Just Symptoms

Fear is the real illness. We are the cure.

That line became the entire foundation of a project I've been building called MannSaathi — an anonymous, AI-powered health companion designed to solve a problem most healthcare apps completely ignore.

The Problem Nobody Talks About

India doesn't lack hospitals. It doesn't lack clinics, diagnostic centers, or telemedicine platforms. The infrastructure is there.

What's missing is the courage to take the first step.

People delay seeking healthcare not because care doesn't exist, but because of things that never show up in a hospital's capacity report:

Fear of diagnosis — not knowing feels safer than knowing
Social stigma — worrying about what people will think
Masculinity norms — "real men don't complain"
Cost concerns — assuming it's unaffordable without checking
Lack of time — life getting in the way
Denial — "it's probably nothing"

So I asked myself a different question: what if the product's job wasn't to fix the system, but to fix the hesitation?

That's what MannSaathi tries to do.

What I Built

MannSaathi is a full-stack, anonymous-first health companion that lets people check symptoms, talk through what they're feeling, and get calm, non-alarming guidance — without ever creating an account.

Core principles baked into the product:

100% Anonymous — no login wall, ever. The moment you ask a hesitant person to sign up, you lose them.
Calm language always — the AI never says "high risk" or "emergency." Risk indicators use amber, not red.
Built for how Indians actually talk — Hindi, English, and Hinglish, not textbook phrasing.
A custom-trained AI model — not a wrapper around ChatGPT or Gemini.

The AI — Trained From Scratch

This was the part I wanted to get right the most.

Most "AI health" projects just call an LLM API and call it a day. I wanted MannSaathi's model to actually understand the way people in India describe symptoms — which is messy, mixed-language, and full of context an off-the-shelf model won't catch.

Base model: xlm-roberta-large
Fine-tuning method: LoRA (Low-Rank Adaptation) — lets you fine-tune a large multilingual model without needing massive GPU resources
Training environment: Kaggle's free T4 GPUs
Output: Detects 22+ symptom categories across Hindi, English, and Hinglish input

The model is hosted on Hugging Face's Inference API, decoupled from the main app, so it scales independently of the frontend/backend.

Tech Stack

Frontend:

Next.js 14 (App Router)
React + TypeScript
Tailwind CSS
Framer Motion for animations
Zustand for state management

Backend:

FastAPI (Python) — handles AI routing and inference orchestration
Prisma ORM
Neon — serverless PostgreSQL

Auth & Infra:

Google OAuth (optional — anonymous mode is the default)
Hugging Face Inference API for AI hosting
Deployed on Render (both frontend and backend)

Design Decisions That Actually Mattered

A few choices that seem small but changed everything:

Dark mode by default — most people check symptoms privately, often at night. Dark mode isn't just aesthetic here, it's functional.
No red for risk levels — red triggers panic, which is the opposite of what a hesitant user needs. I used warm amber instead.
Three-step max flows — every additional decision point is a chance for an anxious user to bounce. The symptom checker never exceeds 3-4 steps.
Anonymous-first, not anonymous-only — registered accounts unlock history and streaks, but nothing meaningful is gated behind login.

What's Next

This is still evolving — next on the list is expanding the symptom taxonomy, adding voice input for users who find typing symptoms difficult, and improving the model's confidence calibration on edge cases.