Seven years in production. Seventeen months instrumented.
Tony Deverill
One engineer. Seven years scaling a UK software company from 30 people to 750. Mechanisms instead of memory.
- 7 years
- inside a UK software company, 30 people to 750Measured 2026-08-16
- 53
- repositories of instrumented outputMeasured 2026-08-16
- 46
- named ways a passing test suite liesMeasured 2026-09-07
- 104 : 91
- checks a machine runs, against decisions written downMeasured 2026-09-07
The engineering story
Seven years inside one of the UK's software companies, then seventeen months of output I can show you — about two hundred days of it fully instrumented. Fifty-three repositories, roughly a million lines of surviving production code and 680,000 lines of tests, effectively none of it typed by hand. Measured machine-wide on 2026-08-16, with generated and vendored content excluded.
In the flagship repository that comes to 104 mechanically enforced rules and gates against 91 written decision records — 48 structural rules, 7 dependency rules, 29 pre-commit gates, 16 pre-push gates and 4 decrease-only ratchets, counted on 2026-09-07. More of that codebase's practice is enforced by machinery than is described in prose. I have not seen a department with that ratio, and not for want of discipline: building a hundred checks is uneconomic beside thirty salaries, and most of them would be rejected as pedantic if a person had to apply them by hand.
None of it was written out of diligence. On a team, implicit knowledge is free to carry and expensive to write down, so nobody writes it. A fresh agent inherits nothing, which makes implicit knowledge unusable rather than merely undocumented — and that inverted the incentive which has defeated documentation for forty years. The standards exist because the alternative was guessing.
The unicorn years
I spent seven years inside one of the UK's software companies, a data integration business I joined at 30 people and contributed across every product and engineering area as it grew to 750 and reached a £1.5bn valuation.
I saw what slows organisations down: process bottlenecks, knowledge trapped in people's heads, operational friction that accumulates over time. I also saw what works and what doesn't when companies try to fix these problems with technology.
Those years built something specific: the ability to recognise failure patterns before they compound, to hold production standards under pressure, and to know the difference between a real engineering decision and a shortcut that creates debt. That is the foundation everything else is built on.
AIntelligent Technologies
We are an engineering practice. We evaluate software from the code outward and build custom projects around your requirements. An evaluation can examine source and contracts, tests and seams, runtime and deployment, browser journeys, crawler responses and authorised security checks. The question and available access determine which checks belong.
A custom build starts from what the software needs to do, who will use it and what it needs to connect to. We write decisions down and run checks automatically as part of the build.
For an evaluation, agree what the written record needs to answer. Ask for the evidence behind a finding and the reasoning behind a proposed change. Keep observed behaviour, findings from the source and questions still open distinct in the record. Discuss a change to the product as its own piece of work.
Registered in England and Wales, company number 16284893. Registered office: 17 Chapel Street, Glossop, England, SK13 8AT.
Values
The principles that shape how we build.
Empowerment, not replacement
AI should help people achieve more with what they have, not take their jobs or remove their agency. That is the design constraint, not a marketing claim.
Build what you use
We use what we build. The agent framework, the evaluation method and the audit that reviews each change run on our own products before they run on anyone else's.
Production discipline
Seven years inside a company that scaled from 30 to 750 taught us what production-grade means. Quality is infrastructure, not a phase.
Mechanism over prose
A rule enforced by memory is not a rule. Every rule that matters here has something behind it that can refuse.
Honest advice
Substance before hype. Where a figure is not measured we do not use it, and where we have corrected one we publish the correction rather than the tidier number.
Practical depth
We go deep on implementation, not broad on buzzwords. Experience with what breaks in production shapes everything we build.
Tony Deverill
Founder
I evaluate software from the code outward and build custom projects around your requirements. The question and available access determine which checks belong in an evaluation.
Bring the decision your team needs to make, or tell me what the new software needs to do. For an evaluation, agree what the written record needs to answer. For a custom build, tell me who will use it and what it needs to connect to.
Get in touch
Questions about software evaluation or custom builds? Tell us about the product or project.