I build and govern data platforms; ML and agents go to production on top of them.

Hourly electricity demand forecasting for PJM, measured on two years of real load from the EIA API and wrapped in the loop that actually matters in production: a scheduled job runs every day, unattended — re-scoring the frozen champion against the actuals that arrived, computing four kinds of drift, deciding whether to retrain, and publishing the numbers back to the repository by pull request. Drift accumulates in public, week after week, and can be pointed at.

• The forecast, measured on real data: 2.72% MAPE against 11.73% for a seasonal-naive baseline on identical folds — 55 walk-forward refits over eight weeks, winning 24 of 24 horizons. The panel behind it is two years of hourly PJM demand with no missing hours, and every artifact derived from it carries "is_real": true.
• The ablation that pays for a design choice: removing the archived day-ahead weather forecasts and refitting costs 41% of the improvement over the baseline. That measurement — not an argument — is what justifies the publication-masked forecast features.

• Python
• LightGBM
• MLflow
• FastAPI
• pandas
• Parquet
• Evidently
• React
• ECharts
• GitHub Actions

The evidence base is stated at its real size: eight summer weeks of real PJM load, and no winter peak has been forecast yet.

A controlled benchmark of Vision Transformers for localising two cervical-spine landmarks in videofluoroscopy frames: nine models by six configurations, 48 of 54 cells filled, a 95% CI on every cell. The dataset is licensed medical imaging and is not in the repository — and the entire analysis verifies without it: no GPU, no data, one command.

• 169 archived training runs collapsed into the grid — with a single traceable run hash behind every number, and a checker that re-derives every published figure from the archive or fails: `aggregate_results.py --check`.
• The finding: freezing the backbone is the most damaging single choice measured, and the best fine-tuned cell reaches 5.18 ± 0.40 px against the 6.31 px baseline.

• Python
• PyTorch
• Hugging Face Transformers
• NumPy
• Matplotlib
• GitHub Actions

The underlying dataset is licensed medical imaging (CC BY-NC-SA) and is not in the repository. The code, the run archive, the figures and the tests are — nothing here reconstructs the data, and CI enforces that.

An end-to-end analytics platform over a monthly mall panel, built for a technical challenge and delivered whole: a medallion pipeline with an ODCS contract, Pandera validation and SHA-256 provenance; reproducible EDA and PCA; four model candidates under identical rolling-origin folds; and the same executive view served twice — a Next.js dashboard and a versioned Power BI project — from the same gold layer.

• Model selection that resists its own marketing: four candidates (Dummy, Bayesian Ridge, GAM, CatBoost) on identical expanding temporal folds; the published champion is a regularised GAM chosen by a parsimony rule — and the repo states that its advantage is not statistically separable, publishing the paired-bootstrap 95% CI of the RMSE difference, [-69; +15] against CatBoost, in the champion metadata.
• The agent never computes a metric. A LangGraph graph scopes the question, picks among allowlisted, validated tools, executes, composes and validates — the LLM writes from the returned payload, with numeric grounding checked, a behavioural evaluation suite and a multi-LLM benchmark in the repo.

• Python
• pandas
• scikit-learn
• CatBoost
• Pandera
• DuckDB
• FastAPI
• Next.js
• LangGraph
• Power BI
• Docker

The panel is one annual cycle of a public 120-record base, so no long-horizon seasonal claim is made; month and event are perfectly confounded, and every figure here is a validation measurement, not a causal effect.

Entity resolution over dirty, overlapping public datasets: blocking to make the comparison tractable, then ML and embedding based matching, ending in a single governed golden record. Designed to run on top of the Open Finance LakeHouse rather than as a separate stack. Not started — this card is a plan, and it says so.

• Blocking strategy first, so the pairwise comparison stays tractable at scale.
• Matching by ML and embeddings rather than hand-tuned string rules alone.

• Python
• Spark
• Embeddings
• Delta Lake
• DuckDB