I build systems
that run█
Ordering intelligence for cannabis retail. AI tooling I own end to end. The owned hardware that runs it all — and the agent that keeps it alive. No mystery platform underneath.
Things that actually run.
Built for real users and real data, on hardware I own. The mess was the starting point every time.
Three Alberta stores under Rocky Mountain Collective used to order off Greenline exports, the AGLC retailer catalogue, and spreadsheet heroics. I built the replacement: an extraction pipeline over ~370,000 rows of order history, an ordering engine with weeks-of-supply logic, category rules, and supplier availability, and a dashboard that turns ordering into a review of one consolidated workbook.
Ordering stopped being spreadsheet trench warfare. The system is in front of real people — demoed to AGLC retail ops in 2026.
Coding agents fall apart on long-running work: context rot, forgotten dead ends, no memory between sessions. I built the fix as infrastructure, not prompt tricks — durable memory tiers, task canvases that tombstone verified dead ends, a JavaScript sandbox for batching tool calls, and subagent lanes that share inference slots so a main session never loses its model.
The agent maintaining this site runs on that harness, on this hardware.
An Open WebUI-style chat platform built to own the whole stack: provider routing, streaming responses, workspaces, prompt modes, tool approvals, and full test coverage. Practical, and not a black box.
What it runs on.
Not a service tour. The four facts that matter.