At TripleH, I lead a team building production multi-agent LLM systems for enterprise clients, from research and prototyping through on-premise and cloud deployment. Our on-premise platform runs agents on open-weight models without external API calls, keeping client data local.
I built and evaluated LLM post-training workflows using SFT, DPO, and RLHF. On a 68-question test, SFT raised a Qwen2.5 model’s score from 15 to 44; I also developed RAG pipelines and LangGraph agents monitored with LangSmith and Arize Phoenix.
My research applies machine learning to soccer analytics, including salary, player performance, and injury prediction. I’m the first author of five papers, and my published models improved prediction accuracy by up to 17% and were deployed as live apps.

