I work at the intersection of responsible AI, data science, mathematics, analytics engineering, and scientific UI/UX. My projects emphasize explicit assumptions, reproducible experiments, strong software architecture, and honest limits—not just attractive outputs. The portfolio shares a versioned Lab Notes identity and research-interface standard: one recognisable family, with discipline-specific accents that communicate rather than decorate.

• Fairness and robustness under distribution shift
• Uncertainty-aware machine learning and model evaluation
• Neural representation geometry and generalization boundaries
• Explainable, accessible research interfaces
• Mathematical algorithms in Rust and WebAssembly
• Data quality, observability, and reproducible analytics
• Planetary dynamics, numerical adequacy, and interactive scientific explanation
• Experimental evolution, multicellularity, and honest causal boundaries
• Accessible science communication — physics, astrophysics, biology, chemistry, and neuroscience, held to the same bar as the AI work: a bounded question, real evidence (public data where possible), and honest limits, not a hot take

An accessible, continuously growing editorial home for the portfolio's research: plain-language articles, a filterable evidence explorer, and a reusable four-step guide for reading scientific results. Every project is presented through the same inspectable structure — question, evidence, finding, and boundary — with direct routes into the underlying interactive laboratory and source code. The production site is live on Cloudflare Workers. Source code

TypeScript · Science Communication · Interactive Learning · Accessibility · Scientific UI/UX

A reproducible Responsible AI laboratory for exploring how distribution shift changes probability calibration, threshold-sensitive decisions, performance, and group-fairness measurements. Visitors can manipulate the population and decision threshold, compare source and target reliability, and inspect uncertainty without installing anything. Version 1.3.0 adds a Robustness Lab — a preregistered synthetic stress study comparing two model families under label noise, measurement error, an unobserved subgroup, and structural misspecification — on top of v1.1's Policy Studio and v1.2's governed external evidence from real 1994 Census data. Source code · v1.3.0 release

Python · TypeScript · Probability Calibration · Distribution Shift · Responsible AI · Scientific UI/UX

An inspectable observability case study built on the European Central Bank's official daily US dollar/euro reference-rate series. Product v1.0.0 separates one prospective live run, a 7,010-prefix replay of the current historical vintage, and a nine-fault synthetic suite. The append-only evidence branch now records source hashes, normalized state and revisions between future runs. At release there is one real prospective run, not longitudinal evidence. Source code · v1.0.0 release

Python · Data Contracts · Observability · SDMX · Public Data · Scientific UI/UX

A reproducible reanalysis of the real, public CHIME/FRB Catalog 1 (536 fast radio bursts, Amiri et al. 2021): does the dispersion measure of repeating bursts actually differ from non-repeating ones, as the original paper's own preregistered test claims it doesn't? The frozen v0.1 study cleanly replicates the paper's pulse-width/bandwidth finding as a validation check, then reports a genuine discrepancy on dispersion measure — traced honestly to sample composition (two prolific nearby repeaters dominating the burst sample) rather than smoothed over. Product v1.0.0 preserves this study. Source code · v1.0.0 release

Python · TypeScript · Astrophysics · Statistical Inference · Real Public Data · Scientific UI/UX

A reproducible test of how well AlphaFold2's per-residue confidence score (pLDDT) predicts real, curated intrinsic disorder. 387 human proteins from DisProt (228,662 residues) joined to real AlphaFold DB predictions by UniProt accession — no model training, just a fresh statistical test of a specific published claim (Alderson et al. 2023, PNAS). The frozen v0.1 study confirms pLDDT is a strong overall signal (43-point median gap, p≈0) but names exactly where it fails: precision collapses on regions that can conditionally fold (6.3% vs. 31% baseline) and on disorder evidenced by HDX-MS, with specific proteins named rather than only aggregate rates. Its hero animation traces one of those named proteins' real per-residue pLDDT values as a wiggling chain, and an interactive threshold explorer lets you drag the pLDDT cutoff and watch precision/recall/F1/MCC update live from the real data, with one click back to the actual preregistered threshold. Product v1.0.0 preserves this study. Source code · v1.0.0 release

Python · TypeScript · Structural Biology · Statistical Inference · Real Public Data · Scientific UI/UX

A reproducible computational-physics laboratory mapping the boundary between quasi-periodic and chaotic motion in the planar gravitational three-body problem, via empirically estimated Lyapunov exponents. Visitors can watch a reference trajectory and a near-identical perturbed twin diverge in real time, export a clip of the divergence, and inspect a frozen, preregistered sweep that tests the figure-eight, Lagrange equilateral, and Euler collinear special solutions against that boundary. The frozen v0.1 study does not claim to solve the three-body problem — it reports what a disclosed numerical method actually finds, including two hypotheses it falsified. Product v1.0.0 preserves this study. Source code · v1.0.0 release

Python · TypeScript · Computational Physics · Chaos Theory · Numerical Integration · Scientific UI/UX

A reproducible measurement of Frankfurt's urban heat island from real DWD station records: the inner-city Frankfurt/Main-Westend station against Frankfurt/Main — DWD's own designated reference counterpart (physically the airport, disclosed rather than glossed over) — across every valid paired day from 1985 to 2025. The frozen v0.1 study finds a measurable but modest gap (+0.455°C, 95% CI [0.432, 0.478], excludes zero) and no statistically significant long-term trend over the 40-year record (p=0.118) — reported plainly rather than reworded into a more dramatic story. Includes an interactive station map computing the real haversine distance and bearing between the two stations from their actual coordinates, no basemap library required. Product v1.0.0 preserves this study. Source code · v1.0.0 release

Python · TypeScript · Climate Data · Statistical Inference · Real Public Data · Scientific UI/UX

A two-stage, protocol-frozen model study of the Io–Europa–Ganymede Laplace angle against 14,610 daily NASA/JPL JUP365 epochs. The original 2001–2030 ablation found the expected model ordering but exposed an unconverged nominal trace. Product v1.0.0 adds a non-overlapping 2031–2040 replication: G4J2 again beat K2, G3, and G4 at a common step, the medium-to-fine trace difference was 0.5735° RMSE, the estimated order was 1.984, and the fine trace reached 0.4699° RMSE against JUP365. Every frozen v1 gate passed. The 2031–2040 values are ephemeris predictions and the integration restarts from JUP365 in 2031—not future observations or an uninterrupted forty-year forecast. The interactive site exposes the orbit, model, and convergence evidence. Source code · v1.0.0 release · Lab Notes article

Python · JavaScript · Planetary Dynamics · Celestial Mechanics · JPL Horizons · Scientific UI/UX

A stable, protocol-frozen research product built from public MuLTEE source data. The historical v0.1 morphology study found positive cell-shape/cluster-size associations in all five anaerobic lines. The v1.0 intervention now shows that engineered tetraploidy increases cluster radius in both tested backgrounds, while the longitudinal comparison rejects tetraploidy as sufficient: at transfer 1,000, mean PA radius is 8.859× mean PM radius despite widespread tetraploidy. The interactive laboratory exposes all 16 chromosome copy numbers, the engineered 2N/4N contrast, and the PA/PM trajectory while keeping measured data separate from explanatory geometry. This is a source-author-data reanalysis, not an independent wet-lab replication or a measured entanglement threshold. Product v1.0.1 clarifies the animation's bonds, steric contacts, and modelled fractures without changing either frozen study. Source code · v1.0.1 release · Lab Notes article

Python · JavaScript · Experimental Evolution · Multicellularity · Reproducibility · Scientific UI/UX

A protocol-frozen test of neural collapse after interpolation on the official writer-disjoint UCI Optical Digits split. Thirty 600-epoch MLP runs compare clean, deterministic long-tail, and 20% symmetric label-noise training against a multinomial logistic reference. Two of four directional gates pass: NC4 remains no worse after zero error in 10/10 clean seeds, and imbalance worsens NC2 in 10/10 paired seeds; continued NC1 and NC2 improvement each reach only 7/10. Clean median held-out accuracy is 96.27%, while the noisy condition is only 81.64% despite a slightly lower median NC2. The Geometry Theatre replays actual saved checkpoints using fixed-basis PCA for display while reporting every NC endpoint in the full nine-dimensional representation. This is one small architecture and dataset—not evidence that collapse universally causes or certifies generalization. Source code · v1.0.0 release · Lab Notes article

Python · Machine Learning · Neural Collapse · Representation Geometry · Reproducibility · Scientific UI/UX

A reproducible cross-dataset test of a fixed P3b endpoint in public EEG data. The electrode, 300–600 ms window, target-minus-standard contrast, artifact threshold, participant-level inference, and stopping rule were frozen before the external amplitudes were inspected. All 13 OpenNeuro participants showed a positive contrast; the mean was +5.65 µV with a 95% confidence interval of [+4.83, +6.48]. The interactive laboratory exposes every participant and both prespecified artifact-threshold sensitivity analyses while keeping the confirmatory endpoint visibly fixed. Product v1.0.0 stabilizes the research product without changing the frozen endpoint or result. Source code · v1.0.0 release

Python · EEG · Neuroscience · Statistical Inference · OpenNeuro · Scientific UI/UX

A transparent known-result reproduction of ORDerly's reaction-condition benchmark. Research product v1.0.0 verifies the primary paper, both versioned Figshare sources, official cleaning logs, and the complete four-variant archive. All four frequency baselines reproduce within 0.46 percentage points of the final peer-reviewed values. An exact audit of 691,142 released reactions finds zero train/test collisions on the declared reactant/product input key and zero exact full-record duplicates across the split. The prespecified v1 audit finds 5.78% canonical product overlap, 80.84% nonempty product-scaffold overlap, and 60.5% sampled maximum product similarity ≥0.70. Neural-model scores remain published references because exact checkpoint/prediction bundles are not in the versioned public release; product overlap is not proof of patent-family leakage or wet-lab failure. Source code · v1.0.0 release · Lab Notes article · Research report

Python · Computational Chemistry · Machine Learning · Data Provenance · Reproducibility · Scientific UI/UX

An interactive Rust/WebAssembly laboratory for understanding what root-finding methods certify and why raw residual alone can mislead. Stable product v1.0.0 preserves the v0.1 solver and v0.2 safeguarded-method studies, then adds five prespecified conditioning cases. Equal forward error coexists with residuals spanning sixteen orders of magnitude, while the repeated-root case reports the simple-root diagnostic as unavailable. All frozen checks pass under the explicitly stated additive perturbation model. The selected teaching suites are not a prevalence estimate, rigorous enclosure, or production-library equivalence test. Source code · Lab Notes article · v1.0.0 release

Rust · WebAssembly · Numerical Analysis · Scientific Computing · Accessibility · Interactive Learning

An interactive cryptography learning application whose cryptographic domain logic is implemented in Rust and delivered through WebAssembly. It combines mathematical explanation, typed error handling, official test vectors, browser tests, accessibility checks, and a documented security model.

Rust · WebAssembly · Cryptography · Accessible Education

Full researched roadmap, including grounded data sources for every future entry: ROADMAP.md.

A second track, editorially distinct from the responsible-AI/data-engineering work above. Every project belongs to the same Lab Notes identity system while retaining a field-specific accent and explanatory visual language. Same standard throughout: every piece states what it contributes, what it found, and what remains unresolved — not just an explanation of settled science. Two formats:

• Explainer builds — an interactive simulation, design, or animation of a real phenomenon currently being discussed or studied, with the underlying model and its limits made explicit.
• Research notes — a short, evidence-grounded, accessible thesis built on public data or literature, testing a specific claim rather than speculating.

The complete research collection is indexed in Lab Notes. Product and study versions remain separate: Jovian Resonance Lab product v1.0.0 preserves the original v0.1 analysis and adds a second, prospectively frozen validation rather than rewriting the historical endpoint. Snowflake Evolution Lab product v1.0.1 preserves the v0.1 morphology study and frozen v1.0 result; its clarified animation separates junctions, steric contacts, and fractures. The study adds a source-compatible genome-duplication intervention and sufficiency evaluation. It does not claim an independent wet-lab replication or a measured genomic–mechanical threshold. Neural Geometry Lab is a separate protocol-frozen v1.0 study: its ten seeds per condition measure algorithmic sensitivity on one fixed writer split, not population uncertainty, and its PCA scene is explicitly separated from full-dimensional NC1–NC4 endpoints.

1. Research questions before dashboards. Every analytical project begins with a falsifiable, bounded question and a protocol appropriate to its design. Known-result reproductions are labelled explicitly rather than presented as blinded or preregistered work.
2. Evidence before claims. Baselines, uncertainty, negative results, and limitations are first-class outputs — falsified hypotheses are reported, not quietly reframed.
3. Sources are verified, not assumed. Every citation and external claim is checked against a primary source before it ships — including in a published paper's own stated methodology, not just its abstract.
4. Architecture proportional to the problem. Clear modules and contracts matter; complexity without evidence does not.
5. Reproducibility by default. Locked dependencies, seeds, tests, CI, citation metadata, versioned releases, and — where the underlying data allows it — a fetch-at-build-time pipeline instead of committing raw third-party data.
6. Responsible delivery. Accessibility (WCAG AA, automatically checked), security, privacy, provenance, bias, and prohibited uses are documented for every release.

Python · Rust · TypeScript · WebAssembly · Machine Learning · Statistics · Statistical Inference · Causal Reasoning · Data Analytics · Testing · CI/CD · Security · Accessibility (WCAG AA) · UI/UX · Technical Writing

Explore the repositories above or open a GitHub discussion on the relevant project. I am especially interested in work that connects rigorous quantitative methods with products people can understand and trust.