Three things, every day. All three are flavors of the same underlying question: how do you make AI actually useful?
I'm interested in AI that does a job end-to-end, not chatbots that answer questions, but agents that take a task, break it into steps, call tools, handle errors, and deliver a result. Voice agents that answer live phone calls. Research agents that synthesize reports. Workflow agents that replace a whole job function.
The research questions I care about: When does an agent work? When does it fail? How do you know before you ship?
Foundation models (Claude, GPT, Gemini) are the engines. The interesting work is the plumbing around them: retrieval architectures, prompt caching, agent orchestration, evaluation pipelines, tool-use error handling.
I've written tens of thousands of tokens of prompts. I've built and broken many agents. My takeaways on all of this are in Prompting for agents and the Patterns section.
When founders and operators ask me to look at their AI architecture, I do. Usually an hour or two, sometimes more. The goal is always the same: give you an honest second opinion. Am I overbuilding? Am I overselling the AI? Is there a simpler path?
If you want that. get in touch.
All three are about the product layer. Not training models. Not pushing benchmarks. Just making AI do real things for real people.