A self-learning, agentic FWA Intelligence Platform combining deterministic rules
with a 3-algorithm ML ensemble — purpose-built for Indian insurers and TPAs at enterprise scale.
Every unpaid rupee starts with an undetected claim
Indian health insurers lose 10–15% of claim payouts to fraud, waste, and abuse annually.
FWA Intelligence detects it before the money leaves.
🔍
Upcoding Detection
Catches hospitals billing for procedures more expensive than what was actually performed — flagged against ICD-10 peer benchmarks.
👻
Phantom Billing
Identifies claims for services never rendered. Cross-verified against discharge summaries using Gemini Vision OCR.
📋
Duplicate Claims
Fuzzy-matches near-identical claims across same patient, hospital, ICD code, and date window — even with minor variations.
🕸️
Fraud Ring Networks
Maps Hospital↔TPA collusion rings with ₹ exposure. Flags coordinated fraud patterns invisible in individual claim review.
📈
Claim Inflation
Tracks average claim amounts per procedure per hospital over time. Flags quarter-on-quarter inflation above threshold.
⏱️
LOS Anomalies
Flags length-of-stay outliers per ICD code — too short or too long — based on statistical percentile benchmarks.
How It Works
From claim submission to fraud verdict in seconds
📥
1. Submit
Claim enters via CSV, API, or TPA portal. All Indian data formats supported.
🧠
2. Detect
13 rules + 6 ML models score the claim instantly. Risk score 0–100 with full explanation.
⚡
3. Act
High-risk claims routed to adjudicators. WhatsApp alerts to TPA. Audit trail written automatically.
Financial Impact
Insurers see the ₹ return not just the flags
The platform auto-computes your fraud savings, ROI, and 6-month forecast daily — directly from live claims data. No spreadsheets. No manual effort.
01
Daily ROI Report
Fraud caught, money saved, false positive rate — auto-computed every day from your live claims. Your CFO sees the exact ₹ impact.
02
6-Month Savings Forecast
Linear regression on your claims trend predicts next month's expected fraud savings. Plan your reserves with confidence.
03
IRDAI Compliance Built-In
Every decision is logged with full audit trail. Compliance scorecard auto-computed. Regulatory calendar auto-updated.
04
Self-Calibrating Accuracy
If false positive rate exceeds 20%, the platform automatically recalibrates its threshold. Gets smarter with every adjudicator decision.
Technology Stack
Production-grade infrastructure built for Indian scale
Every component chosen for reliability, cost-efficiency, and performance at
Indian health insurance data volumes — from claim submission to fraud verdict in milliseconds.
⬡
Google BigQuery
Primary database replacing SQLite entirely. Serverless, scales to billions of rows. 43 tables, every query partitioned by insurer_id for strict multi-tenant isolation.
🐍
Flask + Gunicorn
Blueprint architecture — 4 independent modules (ML, Documents, Intake, Validation). 2 workers × 2 threads. 280s timeout for ML training headroom.
🧠
scikit-learn + SHAP
3-algorithm ensemble on 7 real claim features from BigQuery. SHAP values for per-claim explainability. Auto-retrain on drift detection or 100+ new claims.
☁️
Cloud Run + GCS
Serverless containers, min-instances=1 (zero cold starts). ML models versioned in GCS with backup before every overwrite. Documents at 50GB+ scale.
👁️
Gemini Vision
Document OCR — flash-lite only, no cascade. 20× token reduction via pdfplumber. Hard 100 calls/day cap. 24h cache. Auto-fallback to keyword matching.
🔐
JWT + Secret Manager
HS256 JWT with BQ-backed token blacklist on logout. SECRET_KEY from GCP Secret Manager. 5-layer middleware. Role hierarchy: owner → admin → editor → viewer.
Single HTML file, zero build step. Hash-based routing, 27 pages. 10-second polling for real-time fraud alerts. IST timestamps throughout.
🔗
LiteLLM Gateway
Unified LLM gateway for model swapping. OpenAI for business decisioning, Gemini for OCR, DeepSeek for lightweight inference — all via one interface.
ML Algorithms
Three algorithms. One ensemble verdict.
No single algorithm catches all fraud. Our ensemble combines three complementary approaches —
each blind to the other's weaknesses — with SHAP explainability on every decision.
Ensemble Score Formula
score = IF × 0.50
+ LOF × 0.35
+ HBOS × 0.15
// Auto drift detection
if z_score > 2.0σ → auto_retrain()
Builds random isolation trees. Anomalies isolated faster — shorter path = more anomalous. Best at catching claims unusual across the entire dataset.
LocalOutlierFactor35% weight
Neighborhood Density Analysis
Compares local density vs k-nearest neighbors (k=20). LOF >> 1 = neighborhood much denser — catches hospital-specific patterns rules miss.
HBOS15% weight
Histogram-Based Feature Scoring
Scores each of 7 features independently via histogram bin density. Low bin density = rare value. Catches single-field anomalies invisible to other algorithms.
SHAP ValuesExplainability
Feature Contribution Attribution
Tells adjudicators exactly which feature drove each anomaly score. claim_amount, LOS, hospital_fraud_score, triggered_rules_count — no black-box verdicts.
Union-FindNetwork
Fraud Ring Community Detection
Graph community detection on Hospital↔TPA relationships. Rings with 2+ members flagged with ₹ exposure — persisted to BigQuery.
DBSCANGeospatial
Geographic Fraud Cluster Detection
Density-based clustering on hospital lat/long. Finds geographic fraud zones — surfaced on live Leaflet.js India map with OpenStreetMap tiles.
Tech Vision
Autonomous fraud intelligence that never stops learning
The platform learns from every adjudicator decision, recalibrates autonomously,
and improves its own accuracy without human intervention.
01
Self-Learning Agentic Scheduler
Daily drift check auto-retrains when claim distributions shift beyond 2σ. Weekly hospital risk recalculation. Monthly ICD seasonal refresh. Zero manual effort — the platform gets smarter every day.
02
Multi-Insurer Platform Architecture
Strict BigQuery-level multi-tenancy with cryptographic JWT isolation. Every table, every query filtered by insurer_id. Foundation for insurers, TPAs, hospitals, brokers — one platform.
03
Monetization-First Design
ROI and fraud caught auto-computed daily from live BQ data. Linear regression 6-month savings forecast. False positive auto-calibration. Insurers see exact ₹ return — not just flags.
04
LLM Gateway Architecture
LiteLLM as model-agnostic gateway. OpenAI for business narratives, Gemini for OCR, DeepSeek for lightweight inference — all swappable via config. Vendor lock-in eliminated.
Platform Scope
Built for the entire health insurance ecosystem
From first-notice-of-loss to settlement — every stage of the claims lifecycle,
every stakeholder, every data format Indian TPAs use.