03 — Hybrid AI System · 2025–26
AI Loyalty Fraud Detection
A hybrid AI fraud-detection system for airline loyalty — six detection modules, including an LLM that reasons about whether activity is even geographically possible.
Role
- Sole engineer · 3-month POC
Context
- IndiGo
Year
- 2025–26
Stack
- Async Python, LLMs, MS Graph API, Azure Cosmos DB, Capillary APIs, Navitaire
6
detection modules
0
hardcoded thresholds
3 mo
sole-engineer POC
Overview
As the sole engineer on a three-month proof of concept, I architected and delivered a hybrid AI fraud-detection system for IndiGo's loyalty programme. Six detection modules analyse activity — including LLM-driven geographic-feasibility reasoning that judges whether activity is physically plausible with zero hardcoded thresholds, and self-connecting flight protection.
It integrates the Microsoft Graph API, Azure Cosmos DB, three Capillary APIs and Navitaire. Midway through, I pivoted the architecture from microservices to an async monolith: fewer moving parts, fewer network hops, and far better reliability for production.
What I built
- 01Six detection modules in one hybrid AI system
- 02LLM-driven geographic feasibility reasoning with zero hardcoded thresholds
- 03Self-connecting flight protection
- 04Integrations with MS Graph API, Azure Cosmos DB, 3 Capillary APIs and Navitaire
- 05Pivoted from microservices to an async monolith for production reliability
Architecture
Enterprise work — source is private.
01 Sources
MS Graph API
Azure Cosmos DB
Capillary APIs
× 3
Navitaire
02 Detection
Geographic feasibility
LLM reasoning · no thresholds
Self-connecting flight protection
+ 4 more modules
03 Runtime
Async monolith
Pivoted from microservices
04 Output
Risk findings
Per-module evidence