Notes from the work
Each piece here extends something I shipped: the design decision, the measurement, or the failure that a case study only had room to name. No trend pieces. If it is not grounded in a system I built and ran, it is not here.
Geospatial ML · Data Quality
iNaturalist knows where the trailheads are. A model trained on raw sightings learns the road network, and the corrections are less about clever models than about being honest about who collected the data.
9 min read
Generative AI · Content Operations
On a marketplace that pays per download, an image nobody can find is worth nothing. The pipeline that earned 5,110 downloads starts from a search-volume export, not from a prompt.
9 min read
Simulation · Decision Support
Two strategies, two random fires, and the better strategy loses because it drew the worse wind. Common random numbers are an old trick from simulation statistics, and they are what turn a wildfire model into a decision tool.
9 min read
Geospatial ML · Evaluation
A held-out point a few hundred metres from its training neighbour is not a test. Spatial autocorrelation turns random k-fold into a memory exam, and the fix is to hold out places rather than rows.
10 min read
Generative AI · Game Engines
Video models are unreliable narrators. In a card game where every attack becomes a generated clip, the only safe design is one where the model is never asked a question the rules engine has not already answered.
9 min read
Next: Machine Learning Mushroom Forecasting for Washington State, on .