Case studies
Detailed accounts of how I frame a problem, test the important assumptions, make trade-offs, and respond to what the evidence reveals.
A student research case study that built a leakage-aware benchmark to test whether a person's survey history could be compressed into model weights well enough to predict unseen answers.
What it shows: A good experiment should be able to stop an attractive idea. The useful result was not a digital twin; it was a benchmark, a failure diagnostic, and a better next decision.
An anonymized discovery pilot exploring how structured study context, deterministic checks, and human judgment can make AI-assisted translation review more useful.
What it shows: In a GenAI system, more context is not automatically better. The useful move is to make the decisive context visible, test whether it changes the judgment, and protect exact rules with code.
A note on scope
Most of my professional work comes from confidential, long-running operations. I use that experience to explain how I think without exposing private systems, data or source code.