Our research
The problems we chose to focus on
Alongside the products we ship, we keep a research arm focused on problems that are hard on purpose: those where the state of the art is still moving and the result is worth owning. Our products integrate generative AI; part of this research works one level down, on the mechanisms of generation and reasoning themselves. The list stays deliberately short, worked in depth and under real engineering constraints; the three themes below are where that effort goes today.
Embedded AI acceleration
A traceable, verifiable path from a trained model to FPGA hardware built to be certified for safety-critical, regulated systems, not just made to run fast.
Read the brief Trust & media integrityDetecting AI-generated video
Telling real footage from a convincing fake by the traces generation leaves behind, and staying robust as the generators keep improving.
Read the brief Sovereign sensingRadar for distributed, networked sensing
Radar signal processing built for distributed, networked sensing, so many low-cost sensors can form one shared picture instead of one vulnerable eye.
Read the briefMethod
Three rules the work runs under
Bench before field
Concepts are checked against physics and against real components early, on a bench. Walls are cheaper to hit in a rig than after deployment.
Dated from the first day
Research is recorded and dated as it happens, so anteriority is established from the earliest stage rather than reconstructed later.
Including the dead ends
Research runs against acceptance criteria set before the run, and the outcomes are written up whichever way they fall: the dead ends to the same standard as the successes. The notes below hold both kinds.
Dual use, by design
Where a problem serves civil life and defence alike, we design it as dual-use; where it is squarely a defence capability, we build it as one. Field applicability, operational robustness and ethical constraints are written into the specification with the first prototype.
Research notes
Notes, including the dead ends
Not every line of research becomes a pillar. Some we pushed until they gave a clear answer, including the answer that something does not work and exactly why. We write those up too.
Can a model reason in a language of its own?
Why a model might think in a compact internal code instead of words: the question, the seven designs we explored, and what survived a strict causal test. The overview of the series.
Read the overview Method & resultA reasoning channel must break when you scramble it
The air-gapped architecture, the three-intervention causal-necessity protocol, and the one run it certified: exact numbers, limits and provenance included.
Read the note Honest negativesHow a reasoning channel fails silently
Three of our own dead ends in machine-native reasoning, including the dangerous kind that looks like success on every dashboard while carrying none of the reasoning, and the test that catches it.
Read the noteDiscuss a research problem with us
If one of these problems is yours, or close to it, we are open to focused conversations with partners, integrators and public bodies.
Write to us