I'm a machine learning engineer focused on reliable ML systems, data quality, and evaluation.

I came into machine learning through physics, scientific research, data work, and operations. In my current role, I own a production catalog classification and enrichment pipeline and its staged move from SQL Server to Databricks. I also contribute to reliability work on the larger shared platform.

My recent work includes hybrid scikit-learn and LLM classification for more than 18,000 parts, controlled corrections and backfills across more than 45,000 catalog records, analysis of a dataset with roughly 2.3 million rows, and health checks for about 300 scheduled jobs.

I care a lot about proving that a system behaves correctly before changing it. For the migration, I built automated coverage for the updated read paths and a separate reconciliation validator. That work caught a 34 percent row loss before release. The final backend change has not happened yet. It still needs implementation work and a full development rehearsal.

• Model Release Lab is a full-stack workbench for comparing champion and candidate classifiers, testing release gates, reviewing synthetic failures, and documenting reviewer overrides and rollbacks. View the live app.
• SKU Sleuth is a Streamlit app for testing product classification, quality gates, migration checks, lineage, and safe loading. View the live app.
• Telemetry Anomaly Review is a Streamlit app for finding and explaining unusual points in synthetic spacecraft telemetry. It compares one combined baseline with separate baselines for each operating mode and provides a downloadable review queue. View the live app.
• Weekmark Household Lab is a household planning dashboard with a cash forecast covering 13 weeks, scenario controls, financing checks, synthetic data, and accessible chart alternatives. View the live demo.
• Orbital Mechanics Simulation implements two-body and J2 orbital propagation with NumPy and RK4. It includes quantitative validation, reproducible plots, and the data behind each result. View the live app.

• Define the failure cases and acceptance thresholds up front.
• Keep model output easy to review, trace, and correct.
• Keep measurement separate from the decision policy.
• Make risky changes in stages, with a backup and a tested rollback path.

All public projects use synthetic or illustrative data. They do not contain employer code or private datasets.