AI platform and automation engineer in Toronto. I build agentic AI and automation systems for enterprises, and I measure whether they were worth building.
Three things I care about
- Agentic governance: AI agents writing production code behind spec gates, blast-radius guardrails, and quality gates that make violations mechanical rather than discouraged.
- Automation with receipts: value-stream baselines, honest ROI models, and measured-vs-projected kept strictly separate.
- LLM engineering: per-task model routing, prompt evals calibrated against human labels, and cost engineering. Listenality's enrichment pipeline runs at ~$0.36 per 1,000 tracks because the evals said it could.
Public work (MIT, each with a docs/learning/how-it-works tour for whoever owns it next)
| Repo | What it does |
|---|---|
| evalmine | Scores a model change on your own tasks: pairwise LLM judge with position swap, Cohen's kappa against your labels, schema and execution checks, latency, cost from a pinned price table. Refuses to headline a win-rate its judge cannot defend. Three-tool MCP server so an agent can run the evals mid-task. |
| mcpclerk | A governance proxy for MCP servers: per-tool allowlist (deny by default), human approval for write-class tools, quotas, secret redaction before logging, and a hash-chained audit log that verifies. |
| agentkeel | A framework for shipping production code with AI coding agents: work priced by size, spec-first, four gates with named owners, blast radius bounded by Claude Code hooks. |
| agent-slots | Per-agent isolation for parallel coding agents: one integer derives a worktree, a database, two ports, and a job-queue schema. |
| toilscan | Claude Code plugin that scans Git history for recurring developer toil and recommends the smallest automation that removes it. Every finding cites its commits. |
Background: A decade+ across networking, cloud consulting, and security automation. Azure Solutions Architect Expert, AWS Solutions Architect, Terraform, CCNP.
Toronto · LinkedIn