Independent Applied-AI & Workflow Builder

I build local-first tools, deterministic AI-assisted workflows, and technical systems with an emphasis on inspectability, reproducibility, human review, and honest failure boundaries.

Portfolio · Writing · LinkedIn · Email

Commercial work also includes QuoteForge Local, a white-label quote-calculator package for agencies and local-service implementers.

• Your AI Agent Finished the Task. What Did It Actually Prove? — a practical framework for separating artifact, behavioral, state/provenance, boundary, and independent evidence when AI agents do technical work. The article uses concrete cases where a green or apparently complete state was still narrower than the claim that mattered.

I contribute focused fixes and technical analysis upstream when the work intersects with established projects. These are public evidence artifacts, not blanket endorsements.

• Grid Dynamics Rosetta PR #299 — expanded a dangerous-action guard to catch equivalent git branch force-delete forms while preserving safe commands. A maintainer independently stress-tested the change with a hand-built matrix, randomized differential fuzzing, real Git execution, the repository test suite, and mutation testing before approving and merging it.
• Rosetta PR #319 — hardened dataset lookup ambiguity and removed obsolete authorization behavior after review showed the originally targeted team policy was dead code. Requested code/documentation corrections were addressed, the final head was re-verified and approved, and the PR merged upstream on August 24, 2026.
• Rosetta PR #320 — removed a quadratic dangerous-command matcher path. A maintainer independently pressure-tested the change across millions of inputs, reproduced the performance improvement, requested targeted regressions/invariant comments, then re-verified and approved the corrected head before merge on August 24, 2026.
• Rosetta PR #322 — added focused CLI regression coverage for dataset-name resolution. After review identified two test-fixture gaps, both were corrected and mutation-checked; the contribution was approved and merged on August 24, 2026.
• Super Productivity PR #9619 — keeps section-visible task ordering synchronized with persistent Move Up/Down/To Top/To Bottom actions using the existing reducer architecture. The contribution was merged upstream on August 21, 2026.

• OpenClaw PR #125740 — addressed Skill Workshop routing-description provenance across persistence, revise/apply behavior, legacy migration, and bounded public/model output. Repository-side automated review found no contributor-facing correctness defect while reserving a compatibility-policy decision for maintainers. The PR was closed on August 20, 2026 without merge or recorded human approval, so I do not present it as accepted upstream.

My public corroboration record separates stronger and weaker signals rather than collapsing them into one “validation” bucket. Examples include:

• an external contributor implementing a falsification technique I proposed in their own conformance suite after it exposed vacuous invariants;
• substantive source-level validation and continued technical discussion in CrewAI;
• owner/contributor architectural follow-up in LangGraph, Kimi CLI, and other agent-development threads;
• preliminary local-first curation review of WeaveStudio;
• specialist graph-visualization critique of BuildWorld AI with permission to attribute the feedback publicly.

Review the conservative external-corroboration record

• Problem framing — explicit users, constraints, non-goals, acceptance criteria, and failure boundaries before implementation.
• Systems thinking — dependencies, state transitions, bottlenecks, reversibility, provenance, and evidence boundaries.
• Applied-AI judgment — deterministic logic where determinism is useful; optional model assistance where it adds value; human review where generated output could mislead.
• QA / verification — regression tests, adversarial cases, reproducible issue reports, CI evidence, accessibility checks, and release gates.
• Technical operations — audit records, runbooks, source authority, deployment provenance, exports, SOPs, and implementation documentation.
• Product execution — concept → implementation → validation → deployment → critique → iteration.
• Claim discipline — clear distinctions between implemented, tested, deployed, experimental, externally reviewed, pending, and independently adopted behavior.

Technologies: React · TypeScript · Vite · Next.js · Python · Pydantic · IndexedDB · localStorage · Zod · Vitest · Playwright · Canvas/SVG

Methods: local-first design · deterministic workflows · structured exports · CI verification · regression testing · technical documentation · human-in-the-loop AI · provenance/loss accounting

This is an AI-assisted portfolio. I direct product strategy, requirements, workflow design, scope boundaries, acceptance criteria, verification expectations, source authority, and public claims. AI systems assist with implementation, research, debugging, testing, and drafting; I review, revise, reject, validate, and take responsibility for what is published.

Supporting evidence:

• Portfolio evidence dossier
• Technical decisions
• Recruiter brief
• External corroboration

Commercial availability does not imply verified revenue, customers, active users, purchases, or completed acquisitions. Deterministic scores are heuristics rather than certified predictions. Local-first storage is not automatically encrypted, durable, synchronized, or compliant.

If you work in software, product discovery, developer tools, technical operations, research, AI-agent infrastructure, or local-first systems, the most useful feedback is specific:

• what breaks;
• what is confusing;
• which assumption is wrong;
• where a model creates false confidence;
• what you would remove;
• or what would stop you from using the work in practice.

Follow the public work here on GitHub or start with the five-minute portfolio review path.