I took a career break. I ended up shipping a production AI system.

By Doryan Gowty, Principal, Anneal

I started to write a post saying "hey folks, I built a thing." Then I looked at how much thought had actually gone into it, and a flippant post felt like the wrong container.

I've been off work for two years following a redundancy. It would be easy to say I knocked something out on my break — but that glosses over what the time actually made room for.

Honestly, the redundancy came at the right moment. My son had been given a serious diagnosis and we were moving towns because of it. I spent most of this period focused on him, and tried to make the time off as much about us as I could. Alongside that I had a run of practical projects on the go — finishing some building work on my property, and helping a friend with vintage at their winery.

If you're ever given the opportunity: take the time, and work on something you wouldn't normally.

That last one is the segue. As much as I like wine and had learned a fair bit about it in a previous life, doing vintage was a genuine leap. But it introduced me to a group running a small drinks brand who needed a hand with their eCommerce business — and suddenly I was back on familiar ground.

Four partners, all with full-time jobs, pouring what spare time they had into a side hustle. It's their brand and their operation — I came in to help and wound up building an analytics and automation layer underneath it. Coming in, I asked the question I always start with:

What information do I need to make timely decisions, and what does it currently take to get it?

The answer, at the time, was "a lot of manual work, and not much timely information." Data lived across Shopify, four ad platforms, a 3PL, an inbox, and several spreadsheets. Getting a straight answer to "how did we do this week, and where should the next dollar of ad spend go?" meant an afternoon of stitching. And the customer-service load — refunds, returns, delivery problems — landed on people who already had day jobs.

So over the last while, I built the system I'd have wanted as an operator in that seat. Three parts:

A data and analytics platform that pulls everything into one place — orders, ad spend across Shopify, Meta, Google and TikTok, shipping, and the rest — and turns it into decisions rather than dashboards. The centrepiece is a marketing-mix model that answers the actual question: given what we've seen, where does the next dollar work hardest?

An autonomous customer-service agent that reads inbound email, matches orders against live systems, and resolves refunds, returns, delivery exceptions and address changes end to end — escalating only when it isn't sure. It's been live since late June.

An evaluation framework for that agent — because the moment you let a model take real actions on real customers, "it seemed to work when I tried it" stops being good enough. It tests the agent's behaviour before any change reaches a customer. This is the part most teams building on LLMs never get to, and it's the one I find most interesting.

One thing stuck with me more than anything else, and it's a good preview of where the detail is headed.

Before the agent went live, our testing found it could be talked into showing one customer another customer's order details. Not a model failure — the order-lookup tools simply weren't constrained to whoever was actually in the conversation. The tempting fix is to add "be careful" to the prompt. That isn't a fix, it's a hope. The real fix was architectural: resolve customer identity once, at a single gate, and constrain every downstream action to it. It turns out that LLMs are just too eager to please.

That's the whole job, really. The distance between a demo and something you'd let touch real customers and real money is enormous — and almost none of it is model quality. It's the boring, essential work of deciding what the thing is allowed to do.

I'll unpack a few of these in follow-ups — the platform itself, the analytics behind it, and the evaluation work.

But you don't have to wait for those to see it. I've pulled the whole engagement together as a walkable case study at anneal.io — the agentic workflows, the marketing-mix solver you can actually run, and the engineering discipline behind it. It's the practice I'm building this work under, and this is the first project on it.

Have a look, the deep-dives will follow.