The research side of the work — two anchor areas, medical AI and energy systems, bound by one question: how do we build systems experts can trust and act on? Below: what I'm investigating, what I've found, the open questions, and the papers taking shape. For how it gets built, see Build.
Clinicians don't need another opaque classifier — they need a second opinion they can interrogate. I work on systems that retrieve and present evidence-grounded rationales, so a model's recommendation can be audited against the literature and the patient in front of you.
Grid operators make high-stakes decisions under deep uncertainty. I build probabilistic forecasts that own their error bars, and use synthetic time-series to stress-test them where real data is scarce — so the model's doubt becomes part of the decision, not a hidden footnote.
Four commitments
The goal is a sharper expert, not an absent one. I design for the human who carries the consequences.
If you can't interrogate it, you can't trust it. Explanation is part of the system, not a wrapper around it.
A calibrated maybe beats a confident wrong answer. Doubt should be legible and actionable.
The interesting problems live between fields — statistics, medicine, energy, and human factors.
Working titles — full references and links coming as these mature.
The hard part is rarely the model.
It's the data, the evaluation, and the human who has to use the output. The architecture is the easy 20%.
A metric is a hypothesis.
Every number you optimize encodes a belief about what matters. Choose them like you'll have to defend them.
Talk to the expert early.
The clinician, the operator, the analyst — they will tell you in five minutes what a benchmark hides for months.