TL;DR

I joined a company where the systems already exist, the scale is real, and the problems are different. That’s the point.

What I built

I built the data and AI department at Honeycomb Software from scratch — hired and led ~15 engineers across production RAG systems, vectorisation pipelines, and enterprise AI products. One example: our CV inference pipeline was too slow to run within the SLA. Parts passed uninspected. After targeted optimizations — not a new model, just smarter engineering around the existing one — inference dropped ~40%, and the system caught defects it had been silently missing.

Before that — an ML recommendation platform and MLOps infrastructure from scratch at GR8 Tech. A sports AI system at SoftConstruct that stitched two 4K camera feeds into a panoramic view for real-time broadcasts. No existing platform to extend. No established patterns to follow. Just constraints, a team, and a deadline.

Every system I’ve built, team I’ve led, and constraint I’ve shipped under becomes the lens for the next decision. At some point, the lens stops getting sharper in the same room.

Why the move

Building from scratch teaches you architecture, trade-offs, and how to ship under pressure. But the problems worth solving shift. They stop being “how do I build this?” and become “how do I make this work at scale, inside a living system, with real users, real consequences?”

monday.com is rethinking how work gets delegated to AI agents — not a chatbot bolted onto the side, but agentic AI woven into the product. That’s the challenge I wanted. The codebase carries real history alongside real velocity. Legacy and new code ship side by side, every day.

At this scale, good architecture is table stakes. You also need to sell your ideas across teams and functions, to people who don’t report to you, before you get to build anything.

What I wanted was an environment where product decisions are argued with evidence, not seniority; where the scale is large enough that each decision carries a real cost, and good ideas still have to survive contact with users, metrics, and production reality. That’s the kind of autonomy I wanted to learn from and contribute to.

The honest part

I outgrew the room.

I’m choosing an environment where the feedback loops are faster because the system is already live. Where the constraint isn’t “build it from nothing” but “change it without breaking what millions depend on.” That’s the problem I want now.

Switchbacks, not ladders

Most career advice measures moves on one axis: up or lateral. The axis that matters is whether each role sharpens the lens for the next one.

This move is a shift from a managerial position to an IC role at bigger scale. The best careers aren’t straight lines. They’re switchbacks — the altitude comes from the turns, not the distance.

Leading teams taught me that a system’s real bottleneck is almost never technical — it’s the decision that nobody wants to make. Going back to deeper technical work will make me a better leader the next time around. And after years of context-switching across fifteen people’s work, going deep on one hard problem is a different kind of satisfaction. The craft — algorithms, data structures, performance — matters again in a way it hasn’t for a while.

What stays the same

I still care about systems that survive production and teams that make each other sharper.

MLOps should follow delivery reality, not tool fashion.

The hype says agentic AI replaces engineers — it doesn’t. It adds another floor to the building. The floors below get touched less, but they don’t stop mattering. When the agent fails at 3 AM, someone still needs to understand every layer down to the foundation. I’m now inside a company building this future, figuring out what work gets delegated and what stays human. That makes the question less theoretical.

What’s next

A month in, I shipped my first performance optimization to production. Turns out algorithms and data structures are just as crucial in the frontend as anywhere else — but that’s a story for the next post.

The stage changes. The doctrine doesn’t.