I build inspectable software, then test whether its claims survive contact with reality.
AI-assisted engineering · agent interoperability · local-first products · reproducible verification
View live projects · Inspect external engineering proof · See what is maintained · Contact
| What to evaluate | Direct evidence | What it actually establishes |
|---|---|---|
| Can a change withstand adversarial review? | Rosetta #320 · Engineering case study | A dangerous-command matcher fix merged after independent differential, scaling, and mutation review; not a general security certification |
| Can an observation be separated from an inference? | AgentMeasure DR-005 · Attributed fixture | An external project explicitly credited my limited reviewer input and synthetic fixture; not authorship of the full specification |
| Can an ambiguous tool result be handled conservatively? | CrewAI recovery fixture · Boundary case study | Local, fault-injected SQLite/provider-double behavior; not a live Stripe, distributed exactly-once, or upstream CrewAI guarantee |
Verified upstream record: 9 authored pull requests merged in Rosetta and super-productivity — 7 and 2 respectively, checked October 20, 2026. Rosetta #359 is included in the seven. The separately merged OpenWorldRPG contributions are summarized below and are not part of this count. See the individual PRs, reviewer findings, limitations, and AI-assistance disclosure in the full external-validation record.
- Published article: When Research Evidence Stops Surviving Product Decisions was accepted and published by the ResearchOps Community on Medium on October 22, 2026.
- External repository merges: Rosetta #359 merged after maintainer review. OpenWorldRPG #252 and #253 were merged in a collaborator's repository; #252 reports a UE 5.8.4 build and three focused tests, with its PR documenting remaining limits.
- CrewAI reliability work: Profile-repository PR #40, #41, and #42, plus recovery-fixture #1 and #2, add pinned-runtime dispatch tests, fault injection, SQLite recovery ownership, and evidence-state safeguards. Fixture #1 reports 20/20 deterministic tests passing on its final head. This is mock/local test evidence, not a CrewAI upstream patch, live Stripe validation, or an exactly-once guarantee.
- Additional merged maintenance and review work: Validation Ledger #49 cleared production DOMPurify advisories; Trajectory Fidelity Bridge #46 removed a stale Ruff suppression; QuoteForge #33 and #37 addressed dependency/type compatibility; and Cosmic Atlas #9 added a source-backed reviewer path. Profile and portfolio PRs also refreshed attribution, evidence links, and reviewer navigation.
| Project | Why it exists | Inspect it |
|---|---|---|
| Trajectory Fidelity Bridge | Convert supported Claude Code history to ATIF v1.7 with explicit fidelity/loss reporting, not an invented promise of native session resumption | PyPI 0.4.1 · Verification report |
| Validation Ledger | Trace evidence and counterevidence into reviewable decisions without passing off a heuristic score as truth | Live demo · Release validation |
| WeaveStudio | Build, review, and export local-first workflows with source lineage and human checkpoints | Live demo · Release verification |
| BuildWorld AI | Reproducible graph scenarios for cascades, sensitivity, and systems intuition under explicitly heuristic models | Live simulation · Methodology |
| Cosmic Atlas Scientific | An offline astronomy atlas connecting source-derived catalog data, explicit provenance, and bounded numerical tools | v5.1.0 release · v6 development reviewer guide |
Specify an invariant → build the smallest implementation → introduce a falsifying test → inspect independent feedback → publish only the claim the evidence supports.
My current work includes TypeScript/React applications, Python CLIs, agent-tool boundaries, provenance-aware workflows, and browser/local data models. AI tools assist with research, coding, tests, and drafting; I direct requirements and verification, review changes, and take responsibility for published work.
An upstream merge establishes that a specific contribution was accepted, not employment, organizational endorsement, commercial traction, or approval of unrelated work. Local tests do not establish production behavior or independently verified real-world usage.
- Engineering and security reviewers: start with Rosetta #320 and the recovery case studies, then inspect tests and the upstream discussion.
- Technical operations and product teams: use the five-minute portfolio review and try Validation Ledger or WeaveStudio.
- Agent interoperability reviewers: inspect Trajectory Fidelity Bridge and its compatibility limits.
- Contribution and publication attribution: use the current-work index and source-by-source correction register. This distinguishes authored merges from submitted, unverified, rejected, and incorrectly attributed work.
Technical writing: ResearchOps Community article · Upstream verification · Tool history and ATIF · Verification limits · Reviewer-friendly PRs
Portfolio · Project index · LinkedIn · Email
