I build open-source, deterministic, local-first tools for AI-agent security and governance, repository analysis, and evidence-backed health/training decision support.
Skunky delegates bounded, sanitized work to untrusted LLM workers without exposing the operator's secrets or full project context. Its local broker treats returned artifacts as hostile, enforces disclosure budgets, and records a tamper-evident audit trail.
Threat model · Quick start · Beginner tutorial
| Project | What it demonstrates |
|---|---|
| Skunky | Compartmentalized LLM delegation with hostile-artifact intake and auditable disclosure controls. |
| warden | A zero-dependency Rust policy DSL with a live WASM playground. |
| CapabilityEcho | PR-time detection of new network, subprocess, eval, lifecycle, and workflow-permission signals on exact added lines. |
| recovery-trail | Client-side Apple Health recovery analysis with transparent rule traces and a live demo. |
| fit-ontology | Wearable and intake data unified into an explainable trainer-facing ontology with product screenshots. |
The suite separates policy decisions, runtime enforcement, PR-time detection, transcript review, and consolidated verdicts. There is no LLM in the governance decision path.
decide → enforce → detect → consolidate → observe
Read the full architecture and adoption path.
Browse the complete project catalog, including repository-analysis tools, supply-chain experiments, health/training workflows, and the full agent-governance suite.
- Deterministic first: important verdicts are reproducible and inspectable.
- Local-first when possible: data stays on the operator's machine unless they opt in.
- Evidence-backed: reports expose the inputs, rules, and lines that produced them.
- Conservative health language: decision support, not diagnosis or treatment.
Burbank, California · Rust · TypeScript · Python · Kotlin · React · FastAPI
X @conalhck · Writing · Email