AI automation developer in Toronto. I build AI products for small businesses, and the agent workflow I use to ship them, with tests that prove they're safe to put in front of real customers.
diriye.ca · LinkedIn · Book a call
- Lintel (private): an AI front desk for GTA repair shops. It answers missed calls, quotes from the shop's own price list, books drop-offs and texts the owner. Piloting with two shops; a public case study follows once there are real numbers.
- Barakah Rewards: co-founder of a halal loyalty platform for Muslim-owned GTA merchants, live on iOS and Android. I'm building the owner dashboard.
- Canada Revenue Agency: AI automation developer through the Federal Student Work Experience Program.
| Project | What it does | Status |
|---|---|---|
| Workflow Friction Mapper | A chat agent that reads a workflow you share and returns a grounded brief: friction, bottlenecks and one first move. | Live, open source |
| HarnessGrade | Scores the instruction files around a coding agent (AGENTS.md, Cursor rules, skills) on a deterministic rubric. | Scoring library; CLI and per-rule ablation in progress |
| Hermes Community Skills | A provenance-aware registry for finding and checking community agent skills. | Live |
| Hermes Mobile | The bridge from my retired multi-agent setup. | Retired; why I stopped |
Codex and Claude Code write most of my code. These rules are what make that shippable:
- Cross-model review. Codex reviews Claude's pull requests and Claude reviews Codex's. Nothing merges without passing CI and an independent review.
- Hooks, not hope. Agents can't push to
main, force-push or write env files; the hooks block it and the rules are unit tested. - Code owns the money. A model never sets a price or computes tax. Every model output is schema-validated before use, and customer text is treated as untrusted input.
- Tenant isolation in the database. Row-level security on every table, tested with two tenants.
- Every model call is metered. Tokens, cost, latency and model id go to a ledger.
- Plan, then build. One plan per slice, small PRs, and a
NOW.mdthat separates what's proven from what isn't.
Experiments and build notes, kept separate from finished work: the lab.