I build models on your own data, and the plumbing that keeps them running after handover.
Recent work:
- A machine-learning snowfall forecast for a practising meteorologist: ten winters of numerical weather forecasts trained against a national gridded analysis, two stages (does it snow, then how much), validated walk-forward on eight winters the model had never seen and 18% more accurate than the weather model's own forecast. Delivered as a package he runs himself.
- The asynchronous processing layer for a client's image classifier: Celery, Redis and Docker with retries and live progress, so large batches stopped freezing his interface.
- A geospatial pipeline over OpenStreetMap data that finds buildings by elevation, slope aspect and distance rules, delivered as CSV and a map layer.
- Dashboards that make the output readable: Grafana over a time-series database, and Qlik reports rebuilt where the underlying data feed was broken.
How I work: I start from the question you need answered and the data you actually have, not from a model choice. I report accuracy honestly, measured on data the model never saw, and I say plainly when the data cannot support the question. You get the code, the trained artefacts and instructions to run it yourself.
Tools: Python, gradient boosting, scikit-learn, pandas, Celery, Redis, Docker, PostgreSQL, Grafana.