Small models, built from scratch
Leon models know your business.
Public and private LLMs know a little about almost everything. They generate answers, and sometimes they get them wrong. Leon models don’t just generate an answer. They choose the right answer from what your business knows, applying your rules, your data and your processes, in under a second.
Numbers from the produce model running at try.leonmodel.app. It reads a messy produce name, typos and all, and says which product it is, that it could be more than one, or that it does not know.
Traditional LLMs
Generalists
They understand a little about almost everything and generate the most likely response. Plausible is not the same as right.
Leon models
Specialists
Built around a specific business, its data, rules and decisions, so they select the right outcome rather than simply generate a plausible one. When the answer isn’t in what the business knows, they say so.
How a Leon model learns your business
One framework takes a model from knowing nothing to running a real business decision, and keeps it improving once it is live. Nothing reaches production on a lucky result.
Learn the business
Your products, rules, data and past decisions become what the model knows. Your data stays yours, passed in when needed, never baked into anyone else’s model.
Practise on real cases
The model works through real orders, requests or documents with what it knows frozen. Every decision is logged with the reason it was made.
Review
What it got right, what it got wrong and why, and the cases that were close. The same review a good team lead would run.
Improve
The model retrains on its own reviewed decisions. No borrowed answers, and no guessing outside what the business knows.
Prove it
A new version must beat the current one on cases neither has seen, and pass a fixed test the business signs off. Only then does it go live.
Run and monitor
Live, every answer is timed, logged and watched. The loop keeps running as your business changes, so the model keeps up.
Why games?
Games are where we built and prove the framework. A game gives a clear score, thousands of decisions an hour and no client data at risk, so every change to how our models learn is tested there first, in the open, before it touches a business. The charts below are that proving ground, running right now.
Small on purpose
A model grows only while a bigger one measurably does better. Small models are fast, cheap to run and can sit inside your own systems.
Rewards we understand
Each model is told plainly what good and bad look like for the business. Different approaches run side by side and the results decide.
Every decision traceable
For any answer you can see what it chose, what came second, and why. Nothing is a black box to the people who run it.
The proving ground, live
These models started knowing nothing about their game, and are learning with the same framework we use for business models. After each sprint the current version plays the same 50 test games, and its average goes on the chart. The dashed line is a hand-written scripted player, for comparison.
Loading the latest numbers.
Leon is built by AJQ
AJQ is an information technology consultancy based in Sydney. Leon is how we put small, specialist models into the systems our clients already run: matching, checking and deciding with the business’s own knowledge, in under a second, with every decision traceable.
Every number on this page is measured, and the charts come straight from the running models. When something does not work, the data shows that too.