All work
In development·AI·2026

ClaimDesk

Solo-built AI co-pilot that auto-resolves 80% of warranty claims at 99.4% accuracy. 17 weeks.

ClaimDesk
80.5% · Auto-resolve · synthetic baseline
03 / 05
/ 01
The problem

Self-administered warranty programs drown operators in free-text claims — every one needs facts extracted, policy adjudicated, and a customer email drafted. Most claims are repetitive, but each still eats human time.

/ 02
What I built

Multi-agent LLM pipeline on LangGraph that extracts structured facts, adjudicates against policy with verbatim citations, drafts the customer email, and calibrates confidence via an XGBoost classifier. Auto-resolves at ≥0.70 confidence, otherwise routes to an operator queue with everything pre-filled. Every side-effect is idempotent, every LLM call is traced and cost-tracked, every change is regression-gated against a locked 200-claim eval set.

/ 03
The outcome

Synthetic baseline: 80.5% of claims auto-resolved at 99.4% accuracy, 99.5% verbatim citations, $0.0009 per claim (325× under target), p50 7.0s / p95 9.2s. 100% model stability across self-consistency runs — at the gpt-4o-mini reasoning ceiling. Real-data calibration is the next unlock before production deploy.

Stack
Python 3.12, FastAPI, LangGraph, PostgreSQL, pgvector, XGBoost, Azure OpenAI, LiteLLM + Instructor, Langfuse
← PreviousPropwise TechOperating system for Indian plot & villa developers — live with 5+ developers. WhatsApp-first.Next →ClinicOSMulti-tenant SaaS for Indian private clinics. ABDM-ready, role-aware, AI-augmented.