Now
Small, painful, frequent, and already backed by data or a repeatable workflow.
Founder, Comelse · Production AI operator
I find where AI actually fits—then ship it.
I help operators decide what AI should do, what it should not touch, and whether it can pay for itself. Then my studio builds the useful parts.
Retail platform
100+ showrooms
Production retail workflows spanning AI measurement and factory integrations.
Professional sport
GPS to decisions
Wellness and performance analytics brought into real team workflows.
Agent standard
Tools, not demos
Calendar, Drive, Gmail, browsers, local files—and visible failure modes.
Bring one workflow.
Leave with a decision.
A short working session about the business—not an AI keynote with better lighting. We find the smallest useful move, or decide not to make one.
Later
A useful idea with one missing piece: data, ownership, timing, or a sane success measure.
Never
Demo bait, brand risk, expensive uncertainty—or a human is simply cheaper and better.
What you leave with
- A plain-English fit / no-fit verdict.
- The smallest sensible next step—not a transformation programme.
- The data, risk, and human-approval gaps to resolve before spending.
- If the build earns its place, a route into a Comelse project.
Built,
not briefed.
I ship production systems for clients and run agent experiments on my own stack. That creates a healthy incentive to notice the unglamorous bits: permissions, bad data, retries, handoffs, and whether the thing saves anyone time.
- Production before prototypesLive workflows, observable outcomes, and an owner when something breaks.
- Builder-led judgmentThe person saying “yes” can also scope, ship, and verify the work.
- An opinionated “no”If deterministic code or a competent human wins, the model does not get a participation trophy.
Field notes,
not thought leadership.
Reports from real agent work: short notes when the lesson is small, deeper dispatches when the system deserves showing.
How we turned Hermes into an AI operating system
One general agent became a small operating team with memory boundaries, schedules, permission gates, durable handoffs, and independent QA.
Read the 12-minute dispatch →The agent that rebuilt this website
Git, Cloudflare, DNS, QA, and rollback. Generating the HTML was the easy part.
Read the field note →I put an AI chief of staff on the night shift
Read-only inboxes, trusted code context, isolated worktrees, and a large approval button.
Read the field note →One agent was too easy to trust
Splitting research, building, and verification made failure easier to see.
Read the field note →I let an agent into my inbox
Calendar, Drive, Gmail. It filed a Stripe invoice on my Mac. The useful split was where it failed.
Read the field note →Playing with Hermes
A self-improving agent on a box you own. Useful once it remembers the boring procedures.
Read the field note →DeepSeek can see the screen now
Vendor benches, cheap screenshots, and why API economics matter more than a launch graph.
Read the field note →Have a workflow that might deserve AI?
Bring the ugly version. We will decide whether to automate, simplify, or leave it alone.
Book the session