Open-source portfolio · GERT · EPSILON · The Backtest That Lied · PJM research
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 engineering pull requests | 21 |
| Curated-resource listings | 1, counted separately |
| Upstream repositories | 14 |
| Upstream owners | 13 (11 organizations, 2 individual owners) |
| Historical recipient-repository reach (October 9, 2026) | 253,097 stars |
Contribution status verified October 21, 2026 (UTC): 21 upstream engineering merges and 1 separate directory listing. Own-project PRs and resource listings are excluded from the engineering total and its 14-repository / 13-owner metrics. The star total retains the October 9, 2026 (America/Los_Angeles) snapshot and was not remeasured; repository stars are context, not stars earned by my patches. See the portfolio methodology and dated evidence.
Selected work:
-
StatsForecast #1241: exposed an optional fixed theta in AutoTheta, OptimizedTheta and DynamicOptimizedTheta, with optimizer-aligned validation and fit/forecast regression tests. Merged October 20; published-release availability and adoption are not claimed.
-
SALib #693 — merged a Morris sampling reliability guard for unbounded distributions. The final version validates endpoints before sampling and includes maintainer refinements and regression tests. SALib now accounts for five merged upstream contributions.
-
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 upstream engineering contribution has an evidence-backed writeup covering diagnosis, implementation, validation, review, and lessons learned: explore all merged work →
External resource inclusion: EPSILON's 25-minute Assumption Clinic was included in awesome-quant on October 19, 2026. This is a curated listing, not evidence of learner adoption or validation of trading performance.
| 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.