Enterprise Fraud Detection AI Platform

Problem & Constraints

Fraud cases range from known patterns to adversarial language designed to avoid keywords. Raw case data could not leave the secured environment, yet the system still had to improve processing speed, remain auditable, integrate with multiple client environments, and preserve human accountability.

Engineering Role

Designed and built the three-layer AI architecture, fine-tuned the LLM, created the fraud-knowledge retrieval layer, solved the data-residency constraint, standardized the API, and supported production rollout across enterprise and device-manufacturer clients.

System Architecture

Engineering Principles

• Use deterministic methods where they are faster, cheaper, and more reliable.
• Use LLM reasoning for adversarial cases where traditional methods fail.
• Keep raw data inside the security boundary and preserve human accountability.

Technical Approach

• Fine-tuned Qwen-32B with LoRA and selected r=64, alpha=128 after a rank ablation across r=8–128
• Implemented one-way knowledge distillation and vector export so no raw case data crossed the secured boundary
• Built a fraud behavior-chain knowledge base with FAISS and hybrid semantic/keyword retrieval
• Added human review for high-risk outputs and operational monitoring for production behavior
• Exposed the capability through a standardized API running below 200ms with 99.9% uptime

Production Impact

Reduced case-processing time from 2–3 hours to 20 minutes, reached 85% officer adoption, maintained a complaint rate below 0.1%, scaled to 25+ enterprise and device-manufacturer clients, served 15M API calls per day, and drove 100M+ RMB in cumulative contract value.