Building trustworthy computational tools for real-world decisions.
Reliable machine learning · Scientific open source · Probabilistic systems
Open-source portfolio · GERT · EPSILON
I am a student developer working at the intersection of applied mathematics, machine learning, and scientific software. I build systems that make uncertainty, provenance, assumptions, and failure conditions visible—then test whether their conclusions still hold outside a clean demonstration.
My work has moved from training models and optimizing headline accuracy toward a harder question: when should a computational result be trusted? That question now connects my open-source engineering, energy-risk research, and quantitative decision tools.
| Verified outcome | Snapshot |
|---|---|
| Merged upstream pull requests | 15 |
| Upstream repositories | 13 |
| Open-source organizations | 12 |
| Deduplicated reach of recipient repositories | 246K+ stars |
Repository stars are ecosystem-scale context, not stars earned by my patches. Repositories are counted once; methodology and dated evidence are available in the portfolio methodology.
Selected work:
- MAPIE #953 — added AUROC and AUARC APIs for evaluating uncertainty estimates, including validation, tests, documentation, and review-driven iteration.
- Fairlearn #1674 — integrated confidence-interval detection into the public
MetricFrameplotting API with edge-case coverage. - BeeWare Briefcase #2939 — hardened GitHub Actions through least-privilege permissions, pinned action references, credential-exposure reduction, and zizmor scanning.
- StatsForecast #1175 — added conformal-error prediction intervals and supporting tests.
- NeuralForecast #1563 — implemented FreDF, a loss combining time- and frequency-domain error.
Every merged contribution has an evidence-backed engineering writeup covering diagnosis, implementation, validation, review, and lessons learned: explore all merged work →
| Project | What it investigates |
|---|---|
| GERT — Grid Extreme Risk Toolkit · Live | A probabilistic decision-support system for grid stress, tail risk, scenario intervention, and explicit data/model provenance. Its real-model path is gated by calibration rather than promoted when the evidence is insufficient. |
| EPSILON — Quantitative Decision Lab · Live | A research environment for turning market ideas into falsifiable claims while keeping execution assumptions, provenance, negative evidence, and failure conditions visible. |
| The Backtest That Lied · Live | A forensic ML study in which a near-perfect signal collapsed after leakage repair and failed again under costs and liquidity constraints—leading to a documented decision to terminate the strategy. |
| PJM Extreme-Winter Forecasting | Ongoing research into how ordinary annual validation can conceal probabilistic forecast failure during rare cold events. The manuscript remains in development. |
- Negative results are results. I preserve failed tests and stop hypotheses that do not survive their decision gates.
- Reliability needs more than accuracy. I examine uncertainty, calibration, subgroup behavior, leakage, distribution shift, and reproducibility.
- Claims should remain inspectable. I separate simulated from live data, document provenance, and link public statements to code, tests, or primary records.
- Review is part of the engineering. Maintainer feedback, scope reduction, edge cases, and regression tests are recorded—not hidden from the final story.
I am connecting two lines of work that began separately: applied modeling and reviewed open-source engineering. My next projects revisit real modeling problems with stronger validation, uncertainty quantification, and fairness analysis—using the same standards I learned while contributing to public scientific software.