Writing
Notes from a decade of shipping ML into places where being wrong has a physical cost: on modelling, software architecture, and lately on making coding agents behave like engineers.
Posts
- Jun 2026
Agents follow the rules you actually write down
Coding agents fail the way new hires fail: through missing context, not missing talent. What I have learned writing and maintaining the operating rules of an agent-heavy monorepo.
- Apr 2026
Parse, don’t validate, especially in a data-science repo
A dict[str, Any] crossing your codebase is a unit error waiting for production. Why type discipline pays off faster in ML than anywhere else.
- Feb 2026
When the neural network meets the mass balance
Neural networks extrapolate badly and first principles lack fidelity. How hybrid models cover each other’s weaknesses and earn the right to steer real processes.