Autonomous Wiki Sync
A self-updating wiki that syncs with Notion and source repositories on any change and serves each team in its own format. Built with Claude Code in a 72-hour sprint.
Agentic Automation
The Story
Five years building software, the last three putting language models where real customers meet them. At Dubizzle Labs I led five engineers across CRM, real-time messaging and AI voice. Today, at Hyly.AI, I build autonomous agent automation on MCP, with coding agents as my daily workflow. The through-line hasn't changed: take the system everyone calls a prototype and make it something a business runs on.
- 01
Route permissions, never copy them
On a platform unifying three CRMs, every request is checked by the source CRM's own auth model. Nothing rebuilt, so access stays correct as each product changes.
- 02
Agents need contracts
My four-agent SDLC pipeline works because every handoff has explicit completion criteria. The insight is the decomposition, not the agent count.
- 03
Distrust confident output
The failure mode is an agent that looks like it succeeded. I build evaluation into the loop, like post-call analysis that finds the prompts quietly failing.
- 04
Measure it in production
800ms to 120ms on a reported slow query, traced to one missing composite index. Numbers from running systems, not estimates.