Notes
How the pieces behind Routine Run actually fit together.
December 2025 to October 2026
Structed, milestone by milestone
Structed turns one AI-led client interview into a structured set of requirements. I came in as CTO at the very start and I am still involved. These are the steps it took, in the order they happened.
December 2025
Designed it with the founder from a blank page: the architecture and the agent concept. An outsourced team started on the base.
February 2026
Audited the base, ended the outsourcing and took over all engineering.
March 2026
Traces on every run, model routing, and prompts that domain experts can edit without a developer.
April 2026
Simulated clients: an AI persona can now run an interview end to end, so a change is tested before a real client sees it.
June 2026
The domain taxonomy moved out of code into a database with its own admin, and an MCP server opened interviews to AI assistants.
July 2026
Paid accounts through Stripe, a moderation step before a brief reaches the client, deploys without downtime, and a security pass.
August 2026
Self-serve sign-up with a trial, organizations with seats, and test personas run against the live stack and compared run against run. Selected for the Grishin Robotics accelerator, cohort 4.
17 September 2026
Demo day in London. Marina pitched, and the product could show its own measurements.
October 2026
Still involved.
Along the way: 16 professions and 100+ specializations living as data rather than code, replies down from 20–60 seconds to 1–13, and 1,400+ automated tests.
17 September 2026
Four weeks to a demo day
In late August 2026 Structed joined the Grishin Robotics accelerator, cohort 4, in London: 22 teams, 30 days, then a demo day on 17 September at 22 Bishopsgate, three minutes each on stage.
Marina pitched. Her loop ran every week: pitch, get recorded, get feedback, watch it back, rebuild. Mine was the same loop for the product, as CTO: run an interview, measure it, fix what broke, run it again. Four weeks to turn Structed into something you could check rather than take on trust.
Most of the work was measurement. A question in an interview either lands on something the brief still needs or it does not, so every interview got its hit-rate curve, the admin got a page to compare runs, and the question "did this change help" got answered as a probability instead of a p-value.
Some of it was about not losing what the client said: a brief that quietly dropped facts the client had given, a finished answer lost to a single stray character, and an honest "I can’t answer that" that the agent never wrote down.
Some of it was for the client on the other side of the screen. The interview now tells people how far through they are, and shows them what the agent already knows so they can add to it. The rest was the security review you want done before anyone else looks.
Seventy changes merged between 22 August and demo day. None of them was a slide.
September 2026
Sixty App Intents, one definition
Siri, Shortcuts, Spotlight, widgets, the Action button and Control Center all read the same App Intents. Build them once and six surfaces light up; skip them and none of your app’s actions exist outside the app.
Routine Run has 60+ App Intents. Siri, the Action button and Control Center can start, advance or finish any process, and a step can call out to Shortcuts, Home and Health in the other direction.
That is why this is one piece of work rather than six integrations.
September 2026
Apple’s on-device model writes a routine from one sentence
Writing a routine by hand is the reason most people never write one. So the first screen takes a sentence instead.
Apple’s models turn that sentence into steps, running on the device and on Private Cloud Compute. Nothing goes to a third-party API and there is no account to create.
The same engine then runs it, whether that is an on-call runbook, a shop’s opening checklist or a morning routine.