Senior Principal Data Scientist / AI Engineering Leader
I build high-stakes applied-ML and AI systems that are measurable, auditable and production-ready — from fraud and graph analytics to LLM evaluation and developer-reliability tooling. The common thread across everything below: the evaluation is itself an engineering artifact. Baselines, negative results, uncertainty, cost, leakage and failure modes are reported, not hidden.
| Project | What I owned | Evidence |
|---|---|---|
| ellip2 | AML graph-ML research + discovery architecture | 0.911 ± 0.009 PR-AUC over 121,810 labeled subgraphs; discovery over a 49.3M-cluster background; failed model families and leakage/shift analysis documented; reproducible GPU pipeline |
| relief-probe | Public-data fraud analytics, end to end | 11.4M PPP loans; out-of-time validation against DOJ/SBA-OIG enforcement; 23.8× lift @500 with a bootstrap interval; honest that a simple baseline captures much of the signal |
| Shipi18n | OSS QA architecture + evaluation + upstream validation | 64 repos scanned, 21 shipping a verified broken string, 119 defects verified by hand, 3 fixed upstream in the first week — live tally |
| tsfix | AI code-repair architecture | deterministic + LLM repair layers; real-world benchmark split by workload; < $0.005 per fix |
Role on all four: architect and sole author — problem framing, data and evaluation design, system architecture, implementation, reproducibility, deployment surfaces and documentation.
Findings from my tooling that other projects acted on:
- Solidus — 13 dropped
%{…}interpolations restored in pt-BR; PR merged the same day with three core-team approvals. - Plane — Czech template toasts lost their variables; PR merged in three hours by a co-founder.
- nocodb — my issue on translated/dropped vue-i18n variables led to a maintainer PR that swept all 39 locales.
- Excalidraw — my issue on a
{{max}}→{{mix}}typo led a contributor to fix it and add a placeholder-parity test suite to a 130k-star repo.
Fraud-lead research pipeline over 11.4M public PPP loans, validated against real enforcement outcomes from DOJ/SBA-OIG sources.
This project focuses on turning public records into transparent, defensible investigative leads — not accusations.
Highlights:
- Built a local analytical warehouse from public SBA/DOJ data
- Used anomaly detection and ML experiments to rank potentially suspicious loans
- Added positive-unlabeled learning and LightGBM experiments
- Benchmarked signals against known prosecuted cases
- Included bootstrap confidence intervals and clear model limitations
- Built Streamlit views for analyst-style review
- Added LLM-assisted entity resolution, retrieval, and similar-case workflows
- Framed outputs responsibly as statistical leads, not proof of fraud
Stack: Python · DuckDB · pandas · scikit-learn · LightGBM · Streamlit · LLM workflows · graph/retrieval methods
I use GitHub organizations to separate different kinds of work instead of mixing every project into one personal account.
Open-source QA for i18n locale files — a deterministic linter (missing keys, dropped placeholders, invalid ICU, collapsed plurals across JSON/YAML/ARB/PO/XLIFF/Android/Apple) with an optional, separately benchmarked LLM semantic pass. CLI, core library, MCP server, GitHub Action, Docker, pre-commit.
What makes it more than a linter: I run it on real open-source repositories, verify every finding by hand against the source language, and send the fix upstream — then turn every false positive the scan exposes into a regression fixture. Running tally with links: shipi18n.com/oss.
Reliability tooling for AI-assisted software engineering. Flagship: tsfix — library-aware TypeScript error recovery for LLM-generated code, deterministic quick-fixes first, an opt-in LLM repair layer second, measured on a real-world failure benchmark (98.6% single-file, 40.0% multi-file, 81.4% aggregate, under $0.005 per fix) with the weak cases reported alongside the headline.
I am building public portfolio projects that show the kind of work I can discuss openly when client work cannot be shared in detail.
My emphasis is on:
- Messy real-world data
- Transparent assumptions
- Reproducible pipelines
- Honest validation
- Practical analyst workflows
- LLMs used as workflow support, not magic black boxes
- Small tools that solve specific developer pain points
Python · SQL · DuckDB · pandas · scikit-learn · LightGBM · Streamlit · AWS · LLM APIs · LangChain · LangGraph · JavaScript/TypeScript · GitHub Actions
- Fraud detection and program integrity
- Public-sector and government data systems
- Applied ML that survives honest evaluation
- LLMs for retrieval, triage, entity resolution, and analyst workflows
- Reproducible data pipelines
- Developer tools and workflow automation
- API/spec tooling
- Lightweight SaaS products