For the last six-plus years, my work has circled one question: when an AI system makes a decision, how much can you trust it — and can you measure that trust?
I started asking it during my Ph.D. at Vanderbilt, where I worked on formal verification of neural networks — star-set reachability analysis that gives provable robustness bounds on deep learning models. I kept asking it at MathWorks, building adversarial training pipelines, uncertainty quantification checks, and graph-augmented retrieval for production ML workflows. And I'm asking it now at Princeton IT Services, where I research multi-agent systems for enterprise workflows — agents that route documents, search across departments, and automate approval chains.
The question hasn't changed. The scale has.
What to expect here
I'm committing to a post a week. Topics I plan to cover:
- Agentic AI in practice — what actually breaks when you wire multiple agents together, and how to evaluate agent performance, robustness, and scalability.
- Retrieval done right — RAG pipelines, graph-structured retrieval, and hybrid semantic + behavioral search.
- Trust as an engineering discipline — uncertainty quantification, adversarial robustness, observability, and what formal verification research teaches us about today's LLM systems.
- Notes from the field — papers worth reading, tools worth trying, and lessons from moving between research and industry.
If any of that is your world too, I'd love to hear from you — I'm most reachable on LinkedIn or by email.
See you next week.