I build systems that turn messy evidence into decisions — without losing the path from evidence to answer.
These days, that means making agentic analysis of unstructured data more reliable, controllable, and inspectable at Microsoft.
Before that, I worked on:
- Cloud capacity planning — models and simulation for global infrastructure decisions
- Mechanical reliability — explainable failure prediction and domain-specific language models
- Pharma research — ontologies and knowledge systems for messy scientific information
The domains keep changing. The problem mostly doesn't:
How do you turn ambiguous, incomplete evidence into something useful without hiding how you got there?
I tend to care about:
- agents that can be inspected, not just trusted
- evals that explain failure, not just score output
- abstractions that make fuzzy workflows programmable
- models whose assumptions and provenance survive contact with reality
Outside work, I build things mostly because they annoy me enough not to exist yet — including a wireless split keyboard, its firmware, a shortcut overlay, and occasionally do 3D modeling, simulations and make interesting+practical shapes/furniture from wood.