Deep water

My probation at Apex Lab ended last week. It’s the kind of place where you’re surrounded by sharp, curious people who casually solve hard problems on the way to lunch — I’m glad to be in the ranks. Something about wrapping that period made me sit down and actually look back at the work.

A few months ago I was specced into building a Bedrock Agent with a custom MCP server doing KNN retrieval against OpenSearch. I had used Bedrock once. I had never written an MCP server. The reference architecture I was half-following had been published the week before. I shipped it.

Same quarter: a real-time voice agent for a fleet operator. Driver speaks, model extracts a constraint, system pokes a scheduling API. Nova Sonic, bidirectional streaming, low-latency audio. Hadn’t touched any of that stack. Shipped it.

Right now: a FastAPI service that helps analytics engineers write dbt models against a warehouse with fourteen hundred existing models. Shipping.

That’s the deep water, and none of it would have happened solo three years ago. What changed isn’t me — it’s that I got good at cognitive collab: pair-thinking with Claude Code, which knows the territory I don’t. Claude is the bridge to a domain I haven’t internalised yet; my job is taste, scope, and the question “is this actually right.” We meet in the middle and the work happens there.

Three weeks into a project I’m catching Claude’s mistakes. The real output of working with a strong model isn’t the artifact — it’s the model of the domain that ends up in your head.

Here’s what I wouldn’t have noticed without sitting down to write it: that loop has a shape, and the shape is repeatable. Bedrock, voice, dbt — different stacks, same collab. I’d been doing it three times and it took stopping to look back to see it.

Reflection isn’t a soft skill. It’s the loop that makes the rest of the loops compound. Without it the lessons stay implicit, and implicit lessons don’t transfer to the next project. They get re-learned at full price.

So I’m putting it on a calendar. An hour at the end of every project to write down what actually changed in my head. The exercise costs nothing; the lessons compound.

— Botond