I work on statistical modeling, empirical validation, and data-centric machine learning. My projects emphasize careful experimental design, uncertainty, reproducible code, and clear limits on what the evidence shows.
These links lead to research software and supporting evidence, not a list of peer-reviewed journal publications. The README of each repository explains its data, methods, frozen results, and reproducibility limitations.
| Research project | What the code and evidence demonstrate |
|---|---|
| PISA 2022: model verification and evaluation | Statistical validation using PISA mathematics plausible values, survey weights, Fay–BRR uncertainty, calibration diagnostics, a matched EBM comparison, and unseen-school evaluation. Raw OECD data must be obtained from its original source. |
| Biomedical knowledge-graph reliability | A conditional reliability screen for biomedical link predictions, with a frozen protocol, derived statistics, and external-data/time-sliced validation. It does not establish clinical effectiveness or causal benefit. |
| Bayesian TSK regression and uncertainty | Corrected Bayesian spike-and-slab model averaging, posterior-predictive uncertainty, numerical diagnostics, and tests. The evaluated method is not claimed to beat strong dense-TSK and GP baselines. |
| Tabular-to-image encoding experiments | A controlled comparison across 15 public tabular datasets, exploring categorical features, image encoding, architecture, feature ordering, and conventional tabular baselines. Results are protocol-specific. |
Publication boundary: GitHub repositories, software version tags, continuous-integration checks, and software/dataset DOIs do not indicate that an associated manuscript has been accepted or peer reviewed. Historical journal submissions, where mentioned in a project, are distinct from published papers.
| Software project | What it demonstrates |
|---|---|
| DataSentry | Local-first data-quality detection with evidence-based diagnostics and a reversible propose → preview → repair → verify workflow. |
| Data Science Agent | SQL/statistical analysis and machine-learning workflows with traceable evidence behind generated claims. |
| QuantRisk++ | C++/Python quantitative finance implementations with numerical reference checks and reproducible risk simulations. |
Additional software projects: Interve-AI (interview practice and feedback) · MotionSites (motion-driven frontend experiments).
- Define the estimand or question first: distinguish a demonstrated metric, an interpretation, and a real-world claim.
- Compare against alternatives: use baselines, ablations, sensitivity checks, and acknowledge meaningful negative findings.
- Make results inspectable: publish permitted code, test cases, frozen summaries, data-source instructions, and known reproducibility limits.
- Report scope accurately: CI success tests software behavior; it does not prove general theorems, study novelty, or journal acceptance.
I also contribute to community-maintained projects. For example, upstream PR #12213 is a proposed listing of the DataSentry MCP integration in a third-party community directory, not a project I maintain.