I build local-first AI systems, evaluation tools, and practical engineering products.

The strongest thread across the work is simple: make AI agents cheaper to run, easier to inspect, harder to fool, and safer to trust with real code and real decisions.

• Memento Mori Jester: a local MCP/CLI sidecar that reviews agent plans, commands, diffs, and final claims before risky work ships.
• AgentLedger: a local-first black box recorder for AI coding agents, with command evidence, repo state, audit reports, and handoff bundles.
• TokenSquash: a measurable prompt/reply codec and evidence harness for testing whether repeated agent traffic can be shortened without losing meaning.
• RepoMori: machine-readable repository packs for AI agents and local tools, built around exact source recovery and compact context.
• ManifoldGuard: a reference-bounded output regulator for checking candidate outputs against supplied semantic and relational structure.
• The Gauntlet: a local-first paper and theory stress tester with transparent rule-based verdicts and source-grounded reports.

I am converging the public work into a local AI agent control stack:

• memory and context: RepoMori
• execution evidence: AgentLedger
• safety review: Memento Mori Jester
• token and usage pressure: TokenSquash and Tokometer
• output grounding: ManifoldGuard
• evaluation and stress tests: The Gauntlet, The Marked Bench, and the consequence-agent benchmark work

The private work continues this same direction through AIOS, MiddleOut, and consequence-memory systems.

• Tokometer: a local usage gauge for Codex token burn, rate limits, history, alerts, and exports.
• The Marked Bench: a versioned contradiction-detection benchmark for AI reasoning evaluation.
• Rolefit CV: a local-first CV and job-fit assistant that keeps claims grounded in real evidence.
• ChatP2P: peer-contributed AI compute with signed nodes/jobs, verified results, and credit-based coordination.
• Motion-TimeSpace: a research workspace connecting physics thinking with reproducible computational tools.

• Local-first tools that users can inspect and run themselves.
• Evidence over vague claims.
• Practical release gates, tests, and reproducible demos.
• AI systems that remember what happened, show their work, and admit limits.

• Collaboration on AI agent tooling, benchmarks, safety, observability, and local-first products.
• Technical review of the public tools above.
• Contract or freelance work where reliability, clarity, and shipped artifacts matter.