A model that lives in a notebook saves nobody time. I take ML from data to daily use: the pipeline that produces the training data, the service that serves predictions, and the integration that puts them where your engineers already work.
What I build
- Training-data pipelines: batch generation and validation of labeled data, including parametric sweeps over simulation engines (20k-100k cases per model).
- Serving: FastAPI services in Docker, on Azure ML or containers, with request batching to meet latency targets.
- Integration: predictions inside the tool people already use: a CAD plugin, a web app or an API.
- Validation: checks against a reference tool before scaling, held-out test sets, and a route back to the slow, exact method for inputs the model was never trained on.
How we’d work
- Assessment (1-2 weeks, fixed price): where a model would actually save time, what data you have, and what accuracy is good enough.
- Build: a deployed inference service with tests, documentation and a handover.
Proof
- Production ML at Hoare Lea: models that return results 20-100x faster than running the physics simulation, used by 50+ engineers for 500+ design iterations a month (overheating GNN, daylight CNN, peak solar).
- Property valuation engine: gradient-boosted valuation with calibrated intervals and spatial cross-validation.