Products
Products.
Independent technology Thinkery builds around context and efficient AI systems.
These are our own products, built for problems we kept running into in our own work. Client work stays client work; nothing here comes out of it.
LeanCTX
LeanCTX is a context engine. It decides what a model actually has to read for a given run, reuses what has already been established rather than assembling it again, and ties every answer back to the source it came from.
It sits between the systems that hold information and the model that needs some of it. That position is deliberate: most of the interesting decisions in an AI system are made before the prompt is sent, and they are usually made implicitly by whoever wrote the retrieval code.
It reports what that selection saved, so the effect is measured rather than claimed. You can host it yourself, and it works on small models too — which matters, because a system that only behaves well on the largest available model has not solved the context problem, it has bought around it.
It came out of our own development work and is still the product we use hardest.
LeanCTX proof.
Product proof belongs to the product. Anything shown here is evidence about LeanCTX — how it is used and by whom — and never a statement about Thinkery’s client relationships.
We are not publishing install counts, stars or customer stories here until each figure is current and checked. When they are, they appear in this section, labelled for exactly what they are.
Experiments.
Things we are testing on our own work. Experiment means it runs here first and may not survive; discarded means we learned something and stopped, which we would rather show than quietly delete.
Turns the traces of an AI system into a readable evidence sheet. Built because "why did it answer that?" was taking longer to establish than the answer took to produce.
Memory across sessions, so what was understood once stays available. The open question is what deserves to persist, not how to store it.
A context compressor that justifies what it leaves out. A compressor that cannot explain its omissions is hard to tell apart from one that is losing things.
A fast reader for document piles. It ended up too close to search to be worth maintaining separately, so we stopped.