The Shard Startup Summit wasn’t the end of that day.

Two events, one day

In the morning of 2nd June, we pitched Muesli at The Shard and finished as first runner-up. By evening I was in Canary Wharf, at Revolut's office, for Architecture of Intellect, a session run by the Innovators Guild. A different room in every sense: instead of founders and judges, this one was full of engineers, architects, data scientists, and product people who ship at scale.

Agents that speak IDE

The first talk came from Sergey Ignatov of JetBrains, who works on GoLand and DataGrip. He presented the Agent Client Protocol, an open standard JetBrains built that defines how AI coding agents talk to IDEs. The problem it addresses is easy to state: today, every agent needs a custom integration for every editor, so the whole ecosystem keeps redoing the same plumbing. A shared protocol removes that work once, for everyone. What made the talk worth hearing was that his views on the trade-offs came from building the thing, not from speculating about it.

Where the AI stack gets honest

The second talk was from Ezra Citron, a data scientist at Revolut, on experimentation. His subtitle gave the whole session its spine:

Experimentation: where the AI stack gets honest.

Ezra Citron, data scientist at Revolut, Canary Wharf

That framing has stayed useful. The approach is straightforward: AI features ship behind feature flags, roll out through A/B tests, and get judged on measured impact rather than on how good the demo looks.

Two ideas from his agenda I keep coming back to. First, trusting metrics you didn’t define: most metrics you’ll ever use were built by someone else, and you need a way to decide whether they deserve your trust. Second, peeking without lying: how to look at experiment results early without fooling yourself about what they say.

For someone building a product with AI in it, this is the uncomfortable and necessary part. A demo can make almost anything look useful. An experiment tells you whether it holds up.

Why I keep going to these

An evening like this doesn’t come with a certificate or a trophy, and it doesn’t need to. The value is sitting in a room where practitioners describe how things work at scale, and where they break. I left with notes that apply directly to Muesli, most of them about what we should measure before we claim a feature works.

Originally shared as a short post on LinkedIn.