ExaDev

Case study: UK independent connection provider

One memory for the knowledge an engineering firm runs on.

Our client designs and builds grid connections for renewable generation and storage. It's doing something few firms its size have: turning what its engineers know into a platform it owns. We built it with them. It's grounded in the documents engineers work to, and its AI cites its source or says it doesn't know.

  • Energy infrastructure
  • Built from zero
  • July 2026 to now
  • AI, knowledge graphs
made-up citations in daily testing against the client's own question bank
0
releases to production since July
1,500+
of the code, cloud account and data in the client's name from week one
100%
01

The situation

Our client is an independent connection provider. Its engineers design the high-voltage connections that plug new renewable and storage projects into the grid.

Most engineering contractors buy their software. This one decided to build its own. It's the instinct behind Octopus Energy's Kraken: a platform built to run the business leaner and better than competitors could. Our client is applying the same idea to engineering.

Every design has to meet the right documents, at the right edition, and how to apply them often lives in a few senior engineers' heads. Get that wrong and the result is rework, delay and compliance risk.

02

The hard part

No common format. Every source publishes in its own way, on its own schedule. There's no API.

One document, many names. Engineers refer to the same document in different ways, and editions and amendments pile up on top.

Being wrong is worse than being silent. This is safety-critical work. An assistant that confidently quotes the wrong edition, or invents a clause, does more harm than no assistant at all.

03

Real data from the first week

A demo before the build. Before the engagement started, we built a clickable demo on made-up data, so the people who'd use it had something real to react to before anyone committed.

Then we threw the made-up data away. In the first week we moved the demo into the client's own cloud account and GitHub organisation, removed every piece of synthetic data, and connected it to the engineers' own AI tools, each with their own login.

Decided together, in the room. Early working sessions set the principles: never lock the client into one AI vendor, and let trusted engineers write directly while everyone else, AI included, goes through review.

04

Honest about what it holds

The platform knows which edition is current, what replaced what, and what it does and doesn't hold. When it doesn't have something, it says so.

Tested against the truth, every day. The client built its own question bank, independently of us, recording the right answer before the platform is ever asked. They run it most days over the same connection their engineers use, and file what they find straight onto our tracker.

It isn't perfect. But across every run so far, it has never once invented a citation. When it misses, it says “not found”.

Where a number matters, it's calculated the same way every time and traced back to its source, not estimated by a model. The engineer who signs off a design still owns the numbers in it.

05

AI suggests; people decide

Anything the platform's AI can do, an engineer can also do without AI. When AI makes a suggestion, a person accepts or ignores it. Rules that decide outcomes live as data people can inspect and change, not buried in code.

Every engineering firm has experts whose time is the bottleneck. The platform lets them put what they know to work faster. The expert still reviews and signs off; they just don't start from a blank page.

06

Fast, on rails

Over 1,500 releases to production since July: more than ten a day. Built AI-native by a small team. There's no release manager and no release day: every change that lands is released and deployed automatically, with its version number and release notes worked out from the commits. That pace is only safe because of the rails underneath it.

Every change earns its place. An automated reviewer argues with every change before it merges. A merge queue rebases each one onto the latest code and tests it again before it lands. Mutation testing checks that the tests would actually catch a bug, not just run.

Built to move. Every boundary in the system is a plain contract, so a database, a model provider or a hosting choice can change without rewriting the logic behind it. Being vendor-neutral was the client's requirement, and the architecture enforces it.

Why we engineer with AI, not vibe code
07

What we chose not to build

We didn't build a chatbot over a folder of PDFs, or an assistant that answers from memory. And we didn't make AI the only way in. A person can always reach the same answer, and see where it came from.

08

How we worked

Hire one, get all of us. Joe Mearman, a Lead Engineer, made this client his major focus and carries the context. Around him, the rest of the team rotated in as the work needed them:

  • Francis Williams, our CTO, led the data work and planning for the next phase
  • Johannes Pretorius, a Product Consultant, kept the coverage honest
  • Sophie Denysiuk, an Engineer, brought the web app up to scratch
  • Will Faithfull, our CEO, shaped the plan in the first week and has been in the room for every big decision since
  • Sean Leach and James Jackson, both Lead Engineers, advised from the edges

A log channel both sides write in. Every morning we post what shipped, what's still open and a short “needs you” list in a channel shared with the client. They write back in the same place and file findings straight onto the tracker we work from.

Everything in their name. The code, the cloud account and the data have been the client's from the first week.

09

What comes next

The foundation is in place. The next phase extends the same platform into wider project delivery, so the reasoning behind decisions stays with the company.

10

In their words

“We're proving that the work that we're doing is massively useful, and is well worth continuing to do.”
Founder and CEOUK independent connection provider
11

What we'd tell another founder

You don't have to be a giant to build your own platform. Every mid-market firm runs on a few people whose heads hold the business. That used to be a ceiling. AI has made it affordable to build what they know into a platform the whole company can use. Octopus didn't wait to be a giant before building Kraken, and our client didn't wait either. That's how a young company competes with ones ten times its size.

Tell us what you're building.

You'll talk to one of the senior engineers who run the company, not a salesperson. If we're not the right fit, we'll say so and try to point you to someone who is.