I work the whole path a model takes to a decision: framing the question, building the feature and the baseline it has to beat, designing the experiment that judges it, and then pricing the outcome in money rather than error.
A lower metric isn't a result. So my projects carry pre-registered decision rules, guardrails, confidence intervals, and a written verdict — including the two iterations where the more accurate model lost money and did not ship.
Forecasting and experimentation, drift monitoring, streaming detection, and the pipelines underneath them. Every one runs on real data, states what was measured and what wasn't, and links to its full source. Every card flips — the front is the claim, the back is the evidence.
Does the better forecast actually save money? Tested — and twice refused.
A newsvendor cost simulation and a pre-registered A/B test over the full M5 panel. Two iterations returned HOLD; ordering the cost-optimal quantile instead of the mean is what shipped.
Segmentation treated as a model with a lifecycle, not a one-off notebook.
Windowed RFM features, a six-algorithm bakeoff, and drift thresholds calibrated on the data's own variation, with MLflow champion/challenger retraining behind a Feast store.
A decade of US flight data, cleaned, warehoused, and kept current on its own.
Spark cleans a decade of BTS flights into partitioned Parquet; dbt models it in Snowflake behind 36 tests and a Great Expectations gate. A monthly Airflow DAG ingested a new month unattended.
Live anomaly scoring for three assets on a streaming feature store.
Coinbase WebSocket to Event Hubs, a stateful feature engine, a Redis feature store, and an Isolation-Forest plus autoencoder ensemble. One feature definition serves training and serving.
Tools I've used to ship something on this page — not a list of everything I've read about.
A model that does well in validation is only the start. The questions I find more interesting come after it: does this hold on data it hasn't seen, does it change a decision anyone is actually making, and is it worth what it costs to run?
That shows up throughout this work. A forecasting model is judged by a newsvendor cost simulation and a pre-registered A/B test rather than by WMAPE — and the gate said HOLD twice before it said SHIP. A segmentation model is treated as something with a lifecycle, with drift thresholds calibrated against the data's own month-to-month variation instead of textbook constants, which cut false retrain alerts from 21 windows out of 22 down to the four real regime changes.
I also report what didn't work. Two full iterations went into a model that was more accurate and still lost money, and both are written up next to the version that finally earned its launch. A result you'd bury when it disagrees with you isn't a measurement.
I'm comfortable across the whole workflow — framing the problem, moving the data with Spark, dbt and Airflow, building and tuning the model, wiring the serving layer, and setting up the monitoring that catches drift — on AWS, GCP and Azure.
I'm looking for Data Scientist roles. The fastest way to reach me is email.