← Back to Portfolio
AI for Healthcare

Insurance Pre-Authorization Engine

Turns a doctor's notes into a ready-to-submit insurance form. Checks for missing documents. Predicts approval chances.

Try Live Demo →

The Problem

3-7 hrs

Time from patient arrival to insurance approval. IRDAI mandates 1 hour. Routinely violated.

40%

Of submissions get sent back for missing information. Each round-trip adds 2-6 hours. Patient waits.

15+

Different insurance portals. Each with different forms, different rules, different document requirements.

The bottleneck isn't the insurer. It's the 2-4 hours spent inside the hospital — reading handwritten notes, filling 15 different form templates, gathering reports from 5 departments — before the request even leaves.

Architecture

5-Node Pipeline

Doctor's notes go in. A complete, TPA-ready insurance submission comes out.

Input
Clinical Notes
Doctor's notes (Hindi-English OK)
Lab results + imaging reports
Patient demographics
Input
Insurance Info
Policy details
TPA selection (Medi Assist, FHPL, etc.)
Room category + hospital
Node 1
Read the Doctor's Notes
30 min → 2 min
Parse unstructured notes into structured summary
Assign medical codes (ICD-10)
Extract diagnosis, treatment plan, medical history
Estimate procedure cost by city + hospital
Node 2 — Highest ROI
Check Completeness
40% → 15% send-back rate
Cross-check documents against TPA-specific requirements
Flag missing items before submission
Predict which gaps will trigger a send-back
Medi Assist needs PAC for 55+, FHPL needs cost itemization...
Node 3
Predict Approval
Multi-factor approval probability (0-100%)
Green = submit, Yellow = review, Red = fix first
Shows risk factors + positive factors
Based on diagnosis, cost, policy, TPA history
Node 4
Generate Form
Auto-generates correct TPA-specific form
Standard fields from structured data
Narrative sections drafted by AI
One system replaces 15+ templates
Destination
TPA Portal
Medi Assist
Vidal Health / FHPL
Other IRDAI-registered TPAs
Node 5 — Only if needed
Handle Send-Backs
2-6 hrs → 15 min
AI reads the TPA's query
Finds answer in hospital's existing data
Drafts complete response
Flags items needing doctor input separately
Powered by Sarvam 105B — Native Hindi, Marathi, 22 Indian languages
Key Capabilities

Built for Indian Hospitals

Hindi-English Mixed Notes

'Patient ko 6 months se right knee mein pain hai.' Reads this natively. No translation step needed.

Never Invents Medical Data

If a lab result isn't in the input, it won't appear in the output. Every finding traces to the original notes.

Prevents Send-Backs

Catches missing documents before submission. Eliminates the 2-6 hour query cycle that hits 40% of cases.

Knows each insurer's rules

Medi Assist wants PAC for 55+. FHPL needs cost itemization. Vidal needs conservative treatment proof. All checked automatically.

Gets medical codes right

Right knee vs. left knee = different code. Wrong code = send-back. Cross-references notes for laterality and severity.

Handles pre-existing conditions

Flags them, checks policy waiting periods, adjusts approval prediction. Knows diabetic patients trigger specific insurer queries.

Adjusts costs by city

Same knee replacement: ₹2.5L in Bhilai, ₹5L in Mumbai. Flags costs that deviate from insurer expectations.

Works with messy, incomplete data

Missing vitals, absent lab dates, unclear values. Extracts what's available, flags what's missing by importance.

Learns insurer patterns

Medi Assist queries high blood sugar before elective surgery 80% of the time. System preemptively recommends surgeon's note.

Technical Stack

Built With

AI Engine
Sarvam 105B
Indian AI model. Native Hindi, Marathi, 22 languages. 128K context.
Framework
Next.js 15
App Router, React Server Components, server-side API routes.
Hosting
Vercel Edge
Global CDN, instant deploys, atomic rollbacks.
Database
Supabase
PostgreSQL with row-level security. Typed client.
Type Safety
TypeScript
Strict mode. Every AI response validated against typed schemas.
Security
Rate Limited
10 req/min per IP. API keys server-side only. No client exposure.
PDF Processing
pdf-parse
Server-side extraction. Magic byte validation. 20MB limit.
Output Validation
JSON Schema
Every LLM response validated. Malformed output triggers retry.
Monorepo
Turborepo
Shared packages. Parallel builds. Content-hash caching.
Product Thinking

What I’d Build Next

Auto-Pull Hospital Records

Connect to the hospital system directly. Auto-gather lab reports, scans, and history. The 30 minutes spent collecting documents becomes zero.

Eliminates manual document gathering
Learn from Past Submissions

After 500+ submissions, learn each insurer's real patterns. Medi Assist queries HbA1c 7.0-7.5% at 80% rate but FHPL only at 30%.

Closes the feedback loop
Voice to Insurance Form

Doctor speaks in Hindi-English. AI transcribes, structures, and generates the form. Skips handwritten notes entirely.

Eliminates the paper bottleneck
SARVAM 105B | NEXT.JS 15 | VERCEL | SUPABASE | TYPESCRIPTTry Live Demo →