Gradespace Labs
Gradespace is a digital studio that funds its own R&D. Labs is where that work happens — applied AI, built against real deployments rather than benchmarks, in products people already use. This page sets out the directions we are pursuing and how far along each one is.
Where we actually are
To be clear about where we are: this is an active engineering programme, not a published research record. We have shipped products, we have live users, and we have directions we believe in. What we have not yet done is publish. When we do, it will be listed here.
Directions
These are not finished research programmes. They are the problems that keep surfacing in our own products, written down honestly, with where we have actually got to on each.
We want systems that can be checked, not just trusted. The direction we find most promising pairs a model with an explicit rule layer, so that outputs can be tested against constraints a human wrote down — rather than hoping the model internalised them.
Why it matters
Every regulated buyer we speak to — schools, finance, healthcare — asks the same question: can you show me why it said that? Right now most systems cannot.
Models frozen at training time drift away from the people using them. We are interested in systems that keep updating from live signal without losing what they already knew.
Why it matters
A school in September and the same school in June are different data. Static models quietly get worse and nobody notices until the predictions stop being useful.
Multi-step agents are useful and hard to trust. We are exploring supervisor patterns — a second system whose only job is to check the first one against boundaries it cannot edit.
Why it matters
Agentic tooling is moving faster than the safety patterns around it. The gap between "it worked in the demo" and "I would let it touch client data" is where the real engineering is.
Most organisations will never have internet-scale data. The interesting question is how much you can achieve with a few thousand well-structured examples from one domain, plus good synthetic data.
Why it matters
It is the difference between AI being something a school or an agency can own, and something they can only rent from a company much larger than they are.
Fairness treated as a review step at the end tends to produce a document. Treated as a constraint during the build, it produces different software. We would rather find out which parts of that are true in practice.
Why it matters
Our education work touches young people from very different starting points. A model that is accurate on average and wrong for one group is not a model we would ship.
Partnerships
We are a small commercial studio with live products, real users, and an appetite for the harder problems in them. If you are an academic group, a funder, or an organisation with a problem that looks like one of the directions above, we would like to hear from you.
What we are looking for is a partner who wants applied work grounded in deployment. If you need a published track record on day one, we are not there yet — and we would rather say so now than in month three.
Start a conversationWhat we bring