A software company
and AI lab.
We do two things most companies keep separate: we ship production software,
and we build the AI that powers it.
Two sides. One machine.
Ideas go from a blank screen to live products, fast — and underneath, we're building the models that will run them.
We ship software
Full products, built end to end by a small AI-native team: architecture, code, data, design, deployment. Production apps with real users and real payments — not demos.
We build AI
Language models from scratch — custom PyTorch architecture, our own optimizer and tokenizers, terabyte-scale data pipelines, distributed cloud-GPU training. Our products on our own models, not rented APIs.
The projects.
Everything below is live or in active development — designed, built and operated in-house.
Bkoached
A production AI fitness coaching app — an AI trainer and nutritionist that adapts to each user, running on live AI models plus our own pinned in-house model. Next.js & React Native.
Visit →Pub Night
Real-time multiplayer pub quizzing — players compete live, together, in the same room. Built on our own realtime stack.
Play →BuckshireAI
An AI training platform for UK estate agents — a self-learning assistant plus a live course, bringing AI adoption to a traditional industry. Built end to end in about a week.
Visit →Energy Claims Pipeline
A data product that scores, tags and packages UK company records into ready-to-sell datasets. 11,000+ records and counting.
Learn more →Baller Knowledge Quiz
The UK's first skill-based sports quiz platform — daily quizzes across football and more, with real cash prize pools.
Play →Koachie Zero
Our from-scratch language-model programme: custom architecture, our own training stack, and models grown, trained and served on infrastructure we control.
Inside the lab →Inside the lab.
No frameworks doing the hard part. Every layer of the training stack — architecture, optimizer, tokenizers, data pipelines, serving — written and operated in-house, on a budget that makes cost-per-token a first-class metric.
$ python train.py --model 312m --tokens 40e9 --resume latest [lab] custom architecture .......... loaded [lab] our own optimizer ............ loaded [lab] our own tokenizer ............ loaded [lab] corpus: terabyte-scale, curated in-house [lab] training .... every token measured, budgeted, ours.
Who's behind it.
Steven McKinney — founder. Self-taught, autistic, Glasgow born and based. Before building his own company he worked inside some of the UK's biggest institutions, including as a Fraud Investigator at Santander — where precision isn't optional.
Baller Knowledge runs on a simple idea: one focused person with AI-native tools can move at the pace of a whole team — and the detail-first, systems-in-your-head way an autistic engineer works isn't a limitation. It's the edge.
Institutional discipline, product speed, and a research lab in the basement. That's the company.