Email · LinkedIn · GitHub · X / mrvemzi

Say hello

mrvemzi’s R/W · Writing and work

I build the software around AI models.

I’m an AI tech lead in Google Cloud’s forward deployed engineering team, working across Asia Pacific. I write about production AI, incentives, and decisions under uncertainty. The code is where I test those ideas.

GitHub · LinkedIn · Email


Selected writing

Agents are at Docker. We think we need k8s. We actually need CNCF.

Better agents create harder coordination problems. On standards, control planes, and the institutions that let an ecosystem compose.

Winter Is Coming: The Infrastructure Reality of Production Agents

What sits between a running agent and a reliable one: identity, state, network policy, retries, and the infrastructure nobody wanted to build.

The Asymptote of Good Enough

Latency, price, quality—and the point where another improvement stops changing the outcome.

Tacit Code, Real Friction: The Enterprise Reality Behind AI’s Coding Revolution

Cheap code changes what gets built. Integration, verification, and institutional knowledge still determine what survives.

Anvil: The Inconvenient Remainder of Agentifying APIs

An API can describe its requests perfectly and still leave an agent without the meaning or authority to use it.

LinkedIn Is a Market for Reputational Derivatives

Why professional language becomes easier to applaud as it becomes harder to test.

Notes from LinkedIn

Optimising coding harnesses with AlphaEvolve

Context containment has overhead. I used AlphaEvolve to improve the policy for when Straitjacket should intervene. On the measured named-test path, median latency fell 20.15% and tool-result bytes fell 46.67%. The same thread explains the token economics behind the design.

Building to think

A CLI, a builder API, a compiler: small tools force a problem into a form you can run and inspect. A recent note on why I keep building. Read more on my LinkedIn profile.

The Daily Drop

A proposal for a disappearing fashion lookbook: fewer choices, personal curation, and a brief window of attention.

The Coastline Paradox of Enterprise GenAI

Why closer inspection exposes more uncertainty—and more engineering—inside apparently simple AI systems.


What I’m building