Engineeringintelligentsystems.

Most of my work sits at the intersection of LLMs and infrastructure: fine-tuning transformers with QLoRA, building agentic pipelines in LangGraph, and shipping them as FastAPI services with the evaluation and observability (RAGAS, LangSmith, drift detection) to know when they're actually working.

I care most about the parts most people skip: grounding, failure recovery, and honest evaluation. My agentic systems retry and reformulate instead of failing silently, and I've first-authored research on making models small and fast enough to run on real hardware, not just benchmarks.