I work on reliable LLM systems: systems whose behavior can be measured, inspected, and improved.
My background is a decade of NLP and information extraction, much of it deep-learning-based, judged by external evaluation, including work as a principal investigator on DARPA and IARPA programs at Raytheon BBN. The years since have been at AI startups, leading NLP and LLM work in production, most recently on LLM workflows for clinical documentation and structured extraction.
The subject is now LLMs, but the engineering question is familiar: how do you build a system whose outputs can be trusted?
Start here
For my current positioning on LLM reliability and evaluation, read The agent harness and LLM evaluation, honestly. These series explain how I think about the system around the model: state, tools, gates, traces, labels, judges, regression cases, and eval drift.
For the throughline from my earlier NLP work to current LLM workflows, read Information extraction didn’t disappear. It moved inside the workflow.
The rest is grouped by argument under Series. There is also a list of projects and more about me.