Projects

Selected work, most recent first. Every project was implemented by Claude models under my direction; what each one proves about the method is stated plainly.

Butterfly-collector — governed bug-eradication campaigns

A campaign skill that hunts a codebase with agent fleets, plans fixes, executes them, and generates human dev-run bundles — iterating behind hard gates only a human can pass. Before first use I commissioned a 17-agent adversarial review of the skill itself: 9 defects found (including a consent-bypass on its own hard gate) and fixed before it governed anything real. Campaign 1 ran nine iterations against my own live app; final accounting: 12,449 API calls, ~10.6M output tokens, ~$1,610 API-list-equivalent, with cost-per-finding computed per channel and folded back into doctrine.
Proves: autonomy governance, measurement discipline, willingness to stop on evidence.

Claude TGD — AI triage desktop app

Electron + React + TypeScript, local-first. The thesis: the chat UI is not the product; the LLM triage engine is — it reads captures and proposes item-level entries the user confirms or dismisses. I ran the model evaluation personally (four spike rounds, three model tiers, real handwritten-note captures), disqualified a tier over one corrupted phone digit, then kept testing until the failure root-caused to my prompt design — and canonized the law: models copy values verbatim; deterministic code formats them. In daily use on my real life.
Proves: AI product judgment, trust-boundary design, evaluation rigor.

STATELESS — five-faction RTS, design bible to packaged executable

I authored a five-faction asymmetric design document, directed three complete JavaScript versions in a day, then a from-scratch TypeScript deterministic engine (fixed-point math, lockstep-ready networking) built in five phases under mandatory human checkpoints. Balance came from an AI-vs-AI tournament fleet: an outlier faction pulled from a 92% win rate to 64% over six tuning rounds. Shipped as a Windows executable with hotseat and LAN play. A second, clean-room build from the same document — agent forbidden any project memory — verified the spec alone sufficed.
Proves: spec-writing, phased autonomous delivery under governance, verification by construction.

WhAIte Grimoire — semantic library over MCP

A vector-indexed personal library spanning 25 disciplines (~23,700 discipline-tag entries), embedded locally and served by a Python MCP server exposing seven retrieval tools — deployed across claude.ai, Claude Code, and my own app as a permission-gated connector. My most-used AI tools are ones I built and serve myself.
Proves: knowledge engineering, MCP authorship, local-first infrastructure.

Automatic Job Finder Machine — job-search pipeline

Python pipeline + Flask dashboard + Node docx generation: finds, dedupes, scores, and drafts — and never auto-submits, a trust boundary I set. In genuine use by a real client for live submissions; commissioned Mac port verified 63/63 tests.
Proves: end-to-end product delivery, safety-by-design defaults.

Engine forensics — C&C Renegade

Directed binary forensics that root-caused a 24-year-old desktop gamma corruption via import-table scan (the game imports SetDeviceGammaRamp but never the Get — it structurally cannot restore what it never read) and shipped a double-click fix tool. Follow-on modding plan grounded in empirically proven engine truths that overturned misleading public history — since executed, after the game’s 2025 open-source release, as a complete cel-shaded graphical overhaul from a single one-shot /goal prompt.
Proves: directing deep technical investigation to verified ground truth.

The skills library

Four Claude Code skills — orchestration planning, campaign governance, session handoff, and thought-structuring — each adversarially reviewed or A/B-tested before install, then privacy-scrubbed with an adversarial de-anonymization pass for publication. Browse the repo.
Proves: the habit underneath everything above — learn by shipping, then systematize into durable, versioned capability.