We build AI that checks its own work before it speaks, and computing that gives the same exact answer on every machine, every time.
It writes an invoice with a total that doesn’t add up, names a customer who doesn’t exist, or issues a command a machine can’t safely carry out, all with the same confidence as when it’s right.
That’s fine for a chatbot. It isn’t fine for medicine, finance, law, aviation, robotics, or anywhere a wrong answer costs something.
The industry’s answer to unreliable AI is more: more data, more parameters, more compute. We take a different posture. We start from mathematics, building systems whose checks hold exactly or fail visibly. When something is wrong, the system knows, says so, and names what broke.
How do we make it right more often?
How does it know when it’s wrong?
Where a relation must hold, we check that it holds, rather than hoping a model learned it from enough examples.
A withheld answer with a named reason is worth more than a confident wrong one, and it tells you exactly what to fix.
Results you can reproduce, audit and certify, from the data centre down to the smallest chip.
Can a model check its own work before it speaks?
Every structured part of an answer (amounts, totals, names, dates, commands) is checked against relations that must hold. If any fails, the output is withheld and the failure is named. The system repairs what it can and asks again when it can’t.
Can every machine give the same answer, bit for bit?
Neural networks whose results are identical on any processor and at any scale, with integrity checks running in the background to catch and locate corruption in memory, in transit or in a cached result.
When something breaks, can we fix only what broke?
When an error can be traced to one part of a system, only that part is corrected and nothing else is disturbed. That’s faster, cheaper and more stable than retraining the whole model.
Can a model tell new information from repetition?
Models that give credit for new information rather than the same point said ten times, so they stay sharp on long, repetitive inputs.
Can a model know exactly where every word sits?
An exact way for language models to track position that runs on low-precision integer hardware. In our tests it matched today’s standard within the trained length and outperformed it beyond it.
Can the smallest chips control machines without drift?
A control engine for motors, pumps, displays and actuators that moves smoothly, settles precisely and never drifts, on low-cost chips without floating-point hardware.
Can signal processing in silicon be free of rounding error?
Exact signal rotation, measured on real FPGA hardware, for three-phase power, sensor arrays, robotics and low-power AI chips.
Wrong figures stop at the gate instead of reaching your books.
Failures are named rather than hidden behind a confident tone.
The same input gives the same output, so results can be reproduced and reviewed.
Bit-for-bit identical behaviour on any processor is what regulators and auditors need.
An unsafe command is replaced by a safe one before it ever executes.
Integer-exact computing cuts energy and hardware cost.
We’re building a team that would rather prove an answer than average one. Mathematicians, researchers and engineers who think AI should be held to the standard of the systems it runs.
Algebra, number theory, formal methods. People who would rather prove a property than measure it.
Introduce yourself →Architectures, attention and training that hold up to exact checks.
Introduce yourself →FPGA and ASIC design, integer arithmetic, low-power silicon.
Introduce yourself →Motors, actuators and real-time systems on the smallest chips.
Introduce yourself →Runtimes, compilers and inference engines that give the same answer everywhere.
Introduce yourself →People who find the failure before the customer does, and name it.
Introduce yourself →For partners in regulated and safety-critical industries, hardware teams and investors.
[email protected] →