I build deterministic, local-first tools for governing AI agents: policy engines, MCP runtime enforcement, PR-time drift scanners, transcript review, and evidence-backed reports. I also apply the same design style to repository analysis and conservative health/training decision-support tools.

What I build

• AI-agent governance — policy decisions, runtime MCP enforcement, PR-time drift detection, transcript audits, and consolidated verdicts.
• Repository & supply-chain analysis — codebase orientation, stale-repo autopsies, capability scanning, and dependency provenance.
• Evidence-backed health/training workflows — conservative decision support that exposes inputs, rules, and evidence. Not diagnosis, not treatment.

Pasadena, CA · TypeScript · Rust · Python · React · Node · FastAPI GitHub @Conalh · X @conalhck · dev.to/conalh · conal.hg@gmail.com Open to AI-agent infrastructure/safety, developer-tooling, and full-stack roles.

A local-first stack with one job per tool and one shared schema, so the pieces compose instead of overlap: decide → enforce → detect → consolidate → observe. There is no LLM in the decision path.

Full walkthrough: AGENT_GOVERNANCE_STACK.md — the whole suite end to end, with a diagram, a failure-mode map, and an adoption path.

flowchart TB
    subgraph runtime["Runtime · on the MCP wire"]
      direction LR
      warden["warden — decide<br/>allow · deny · ask"] --> barbican["barbican — enforce<br/>before the call reaches the server"]
    end
    subgraph detect["PR time · detect"]
      ScopeTrail
      PolicyMesh
      CapabilityEcho
      TaskBound
      SessionTrail
    end
    detect --> GovVerdict["GovVerdict — consolidate<br/>one PR verdict"]
    barbican -.->|audit findings| GovVerdict
    AgentPulse["AgentPulse — observe<br/>live session trajectory"]
    core["agent-gov-core — one Finding schema, shared by every tool above"]

    classDef hl fill:#0c4a6e,stroke:#0369a1,color:#e0f2fe
    class warden,barbican,GovVerdict,AgentPulse,core hl

Field-tested. Beyond the synthetic agent-gov-demo, I ran the whole stack against a real open-source background-agent coding platform I did not write — runtime MCP enforcement, credential-broker authorization, and a pre-PR capability gate — and fixed a cross-component bug it surfaced. Anonymized writeup: agent-gov-fieldtest.

Standalone tools for understanding codebases, reviewing risky changes, finding documentation drift, and verifying dependency provenance.

• repo-brief — orientation layer for unfamiliar repos: architecture map, key files, hotspots, run commands, and where to start.
• project-autopsy — evidence-backed autopsy reports for stale repositories, over a deterministic, CI-tested core; CLI plus a Next.js report UI.
• docs-debt-radar — scans repositories for stale, missing, and drifting documentation claims.
• overreach — Rust capability scanner for diffs, files, and repos: network calls, subprocesses, sensitive-file reads, curl | sh, disabled TLS, hardcoded secrets.
• tofulock — Go. Locks and verifies Terraform/OpenTofu module sources by commit digest.
• cpan-integ — Perl. Consumer-side, install-time artifact-hash verification for CPAN distributions. Experimental.
• timecal — cross-agent time-calibration corpus served over MCP, countering the engineer-weeks prior agents inherit. On PyPI.

Conservative decision-support tools — not diagnosis, not treatment recommendation. Each one exposes its inputs, the rules that fired, confidence limits, and the raw evidence, so a human stays in the loop on every call.

• fit-ontology — trainer-facing client intelligence: unifies wearables, intake, and ACSM guidelines into a queryable model with explainable rules traceable back to the metric rows that fired them.
• recovery-trail — athlete-facing recovery briefing from an Apple Health export. Runs 100% client-side; shows HRV, RHR, sleep, load, ACSM-aligned verdicts, and rule traces. Live demo.
• nutrition-experiment-lab — personal n-of-1 nutrition experiment notebook with adherence tracking, confounder notes, confidence, and transparent next-test suggestions.
• injury-return-to-play-tracker — clinician- and coach-facing workflow for phase progress, functional-test evidence, workload tolerance, and human clearance decisions.
• academic-load-burnout-monitor — student workload planner with explainable pressure signals, check-ins, and recovery-aware next actions.

• Deterministic first — the important decisions are reproducible and inspectable, not model-dependent.
• Local-only when possible — tools run on your machine; no data leaves it unless you opt in.
• Evidence-backed reports — every verdict traces back to the inputs, rules, and lines that produced it.
• No LLM in governance decision paths unless explicitly opt-in.