I like talking about the parts that get left out.
Getting AI systems to behave inside enterprises that can't afford them misbehaving. I speak about both halves: building with agents, and building agents that ship.
Most recent talk
The slow part was never the typing
Plenary session, Globe DevCon 2026
Ten engineers took a B2B enterprise platform from zero to first release in 108 days, roughly 40% ahead of a six-month plan, with an agent drafting at every stage.
The talk is the operating model that made that safe rather than reckless, the three lessons that cost the most, and where the bottleneck moves once building gets fast.
Topics
Steal this: managing your AI teammate
Seven rules, none of which are AI tricks.
Gate what you can’t undo, not every step. No claim without evidence. Keep changes small and the process smaller. The workshop version of the keynote.
For engineering leads adopting agents without losing control of quality.
The model can’t file anything
Agentic AI that is structurally incapable of doing the wrong thing.
Most demos hand the model the keys. In a regulated enterprise you can’t. Splitting the agent from the credential, putting confirmation in code, and what it costs to run. Includes the failures.
For engineers putting LLM features near a system of record.
Guardrails that aren’t just a system prompt
Governance written as code instead of instructions.
Capability, identity and grounding guardrails — and the part people skip: what trips, what fails closed, and what gets replaced before the user sees it.
For teams told to make an AI feature compliant, unsure what that means in code.
Thirty seconds was the symptom
One cause, three numbers, and why fixing the obvious layer wouldn’t have held.
A notification service answering in thirty seconds, failing a third of the time, deadlocking in production. Observability data down to lock contention. Ends at ~318ms and zero errors.
For anyone who has inherited a system people have stopped expecting much from.
Rollback is a design decision
Moving 1.6 million records live, and never needing the brave plan.
Arranging a migration so that going wrong costs you nothing. Source data left intact, deltas synced behind the batch, rollback reduced to reverting code.
For anyone facing a migration they can’t take downtime for.
Shipping to four markets at once
Regional delivery, localization, and the coordination nobody budgets for.
Rolling features across countries where requirements, regulators and timelines don’t line up. Mostly about sequencing, with the technical parts underneath where they belong.
For teams going multi-market for the first time.
Formats
Conference and meetup talks, 20 to 45 minutes. Internal brown-bags and deep dives. Hands-on workshops. Podcasts, panels, and AMAs with junior engineers.
Where this comes from
I led development on the platform in that keynote and run skills-development programmes for engineers. Everything here comes from systems that are running or in acceptance testing.
Working on the same thing?
Happy to share the useful version, including the parts that didn't work.