Verify through the surface people use.
A 200 response carrying nothing is not a working feature. Open the browser, read the payload, compare the number against one computed somewhere else.
Head of Technology and Architecture at CODVO.AI
I build AI systemsthat hold up on aMonday morning.
Agents, geospatial intelligence, meeting intelligence and the infrastructure underneath them. Built for industrial and enterprise teams, then run in production where the answers have to be right.
- projects tracked
- 111
- active this week
- 13
- agent sessions
- 810+
As of October 2026
About
Head of Technology and Architecture at CODVO.AI.
I lead technology and architecture at CODVO.AI, where we build applied AI for critical operations. My work runs from the first prototype to the system that carries real load: agents, maps, meetings, machines, and the platform they all sit on.
Most of it is for industrial and enterprise teams in energy, manufacturing, healthcare and data centers, where a wrong answer costs real money. So I care less about the demo than about the Monday after it.
Work
Six kinds of problem, one way of finishing them.
Agentic systems
CODVO.AI productNeIO
Agents that plan, act and check their own work inside real workflows, with evaluation harnesses that decide whether a change ships.
Geospatial intelligence
CODVO.AI productAgenticGIS
Plain-language questions over maps and asset data, answered against real geometry so the count and the pins agree.
Meeting intelligence
Conversations turned into decisions, owners and follow-ups, with every line traceable to the moment it was said.
Industrial and predictive AI
Early warning for rotating equipment and plant systems, trained on data scaled to how the machines really behave.
Video and audio AI
Annotation and analysis pipelines for surgical video and speech, from capture through review.
Platform and infrastructure
The part nobody demos: cloud, containers, release gates and a private network of machines that keeps all of the above running.
Method
Rules I work by, each one paid for by a real mistake.
Derive, do not invent.
Demo data has to be internally consistent and anchored to real magnitudes. A domain expert spots the lie in one glance, and everything after it loses credibility.
Design the loop, not the step.
Agents build, other agents review, and a person comes in for the hard parts. Every loop has a cap and a check before anyone says done.
Production writes get eyes-open approval.
Speed is welcome everywhere except the one place a mistake cannot be undone. Incident pressure does not waive the confirm step.
- TypeScript
- Next.js
- React
- Python
- FastAPI
- PostgreSQL
- Supabase
- Azure
- Docker
Contact