Software engineer building trustworthy, reproducible, and inspectable systems.
I focus on the engineering boundary between AI capability and operational trust: evidence, provenance, deterministic evaluation, privacy-aware debugging, and failure-resistant developer tools.
| Project | What it demonstrates | Status |
|---|---|---|
| DiffProof | Portable, privacy-safe, verifiable change evidence capsules for pull requests with impact findings, deterministic integrity, offline verification, SARIF/JUnit-ready CI output, and a composite GitHub Action | v0.1.0 |
| FaultPack | Portable, privacy-first, verifiable failure evidence with safe capture, redaction, integrity checks, replay, reduction, differential comparison, Ed25519 signatures, CI reports, and a composite GitHub Action | v1.0.0 Release |
FaultPack is the clearest expression of my engineering approach: small stable contracts, fail-closed behavior, deterministic evidence, and explicit security boundaries. It is local-first and does not require a hosted account, model call, or implicit upload. The v1.0.0 release includes a tested wheel, source distribution, and architecture documentation.
| Signal | Current focus |
|---|---|
| Reliability | Reproducible failure evidence, regression fixtures, and deterministic replay |
| Trust | Provenance, signatures, privacy boundaries, and explainable verification |
| Developer experience | CLI-first workflows, GitHub Actions, SARIF/JUnit reports, and maintainable contracts |
| Project | What it demonstrates |
|---|---|
| CorpusSeal | Evidence-first benchmark contamination and dataset integrity auditing with deterministic exact/near matching, SARIF, HTML, and GitHub Actions |
| BidiFence | Deterministic RTL/i18n conformance checks for Playwright with SARIF, baselines, and Arabic fixtures |
| TraceSift | Offline causal diagnosis and privacy-safe regression fixtures for AI-agent traces |
| VeriTrace | Deterministic conformance and replay testing for agent governance |
| ML ProofLedger | Portable, verifiable evidence manifests for machine-learning runs |
| Mizan | Evidence-backed Arabic claim verification with abstention and reproducible evaluation |
AI reliability and observability, benchmark and dataset integrity, OpenTelemetry-compatible trace contracts, reproducible ML evaluation, privacy-preserving artifacts, policy-as-code, Python tooling, API design, test architecture, and open-source maintenance.
Python · TypeScript · FastAPI · pytest · GitHub Actions · OpenTelemetry · Docker · PostgreSQL · React
I prefer small stable contracts over opaque integrations, fail-closed behavior over optimistic guesses, local-first workflows where sensitive data is involved, and documentation that states limitations as clearly as capabilities.
The best way to collaborate is through GitHub Issues and Discussions on the relevant project.