Deep · daily work since 2023
Time-series ML & anomaly detection
Finding the moment something goes wrong in multi-sensor streams, and proving the detector is right before anyone relies on it.
- Streaming anomaly detection Production Causal detectors, statistical limits, live and recorded data
- PyTorch · LSTM, autoencoders, VAE Applied Anomaly models 2024–25, ONNX export
- scikit-learn · k-means, Isolation Forest Applied Outlier detection in the comparison tool
- Evaluation design Deep Threshold sweeps, leave-one-out validation, ablations
- DSP · FFT, filters, spectrograms Applied Signal studies and performance pages
- Statistics · Kalman filters, Gaussian processes Deep Statistical limits, PhD
Most of my industrial work is time series: load, vibration and strain from sensorized cutting tools, sampled fast and full of things that look like faults but aren't. The skill is less about the model and more about the evaluation around it: leave-one-out validation that matches how the detector will be used, thresholds chosen where the result is stable, and ablations that are allowed to say "no, more data made it worse".
I am comfortable with deep models and have shipped them, and I am equally comfortable choosing a three-line rule when the evidence says it is better.
Where it shows up
- 2026 A fault detector the shop floor can trust A causal streaming detector for cutting-insert faults, taken from a Python study to shipped C#. It caught every labelled fault without a single false alarm.
- 2025–26 Comparing a cut with a good one, while it happens A live view that streams machining data against a known-good reference curve, so an operator sees a deviation during the cut instead of in the report afterwards.
- 2024 Deep models inside a .NET product LSTM, autoencoder and VAE models trained in PyTorch, exported to ONNX and run on live sensor streams inside the product, with latency monitored against the budget.
- 2026 Machine learning, from first principles to production A curriculum of about fifty lessons written for colleagues, from autograd built by hand to GPT, LoRA, RAG and MLOps, alongside a Python course I wrote and taught.