At BrioAI, I built BrioDocs, a private document intelligence platform for legal professionals that runs LLMs locally so confidential material stays in the user’s environment. I designed multi-step LangGraph retrieval workflows and citation validation against source documents and legal authorities.
I also optimized local inference with llama.cpp and quantized Qwen models, reaching 20 tokens per second on budget laptop hardware. I instrumented model upgrades to compare response quality, grounding, and latency.
At Intelygenz, I developed an agentic loan-document auditing system and productionized a real-time fraud detection platform handling 300+ requests per second on AWS EKS. Earlier, at Veridas, I developed biometric verification and presentation attack detection models evaluated through NIST benchmarks, and optimized CPU inference latency by more than 20% with OpenVINO.
