Writing · evidence · decisions
What the work taught me, not what the internet already summarized.
Each essay starts with a real project and an inspectable artifact. External references add context, while claims about my work stay bound to the same public evidence and publication limits as the case studies.
01 · Forecasting · calibration · evaluation
When should you trust a probabilistic forecast?
A practical trust test for probabilistic forecasts: proper scoring, calibration, complete coverage, leakage-resistant evaluation, reproducibility and published failure modes.
02 · Computer vision · metrics · validation
Why accuracy alone is not enough for oil-spill detection
A metric-design case study from Sentinel-1 SAR segmentation: why rare oil pixels, look-alikes and deployment domain gaps make overall accuracy a weak headline measure.
03 · AI agents · provenance · governance
What should an AI agent do when the evidence is missing?
A practical governance pattern for agentic systems: preserve unknowns, trace material claims to evidence and separate content generation from authority to take external action.