My AI/ML work is validation-led: less about chasing state-of-the-art benchmarks, more about making machine learning systems that can be understood, challenged, and trusted where they matter. An ML model that can't be validated is a liability wearing a performance metric.
ML validation toolkit
Validating ML models means checking the data before the algorithm: leakage between train and test, distribution shift between development and production, unstable features, and evaluation setups that flatter the model. I build tooling that makes these checks routine rather than heroic — see the skeptic's validation playbook research note and the cardiac surgery risk model case study.
Governance for AI systems
As AI systems take on consequential decisions, they inherit the governance expectations long applied to traditional models — conceptual soundness review, independent validation, change control, and ongoing monitoring — adapted for systems that learn, drift, and behave emergently. I write about that mapping in the LLM evaluation harness.
Reproducibility
An experiment you can't reproduce is an anecdote. Versioned data, seeded randomness, pinned environments, logged hyperparameters — the unglamorous plumbing that turns ML work from alchemy into engineering. Reproducibility isn't overhead; it's what makes validation possible at all.
Related projects
- Clinical Risk Predictive Model — a validation-heavy ML case study: stacked ensembles, R² skepticism, residual bias analysis