Company · AI lab
An AI lab that shows its limits.
Alelyon is an AI lab building models that run, learn and act on hardware people already own. We are developing our own line of models and the stack around them: a GPU compute kit on the open Vulkan standard, sub-bit compression for large models, Angel, a private desktop assistant, a simulator for training models, a chat workspace, the Nexus workspace for models and agents, and a verification pipeline with a public checker for applicable replay under independently obtained inputs and a public key pinned out of band.
Operating principles
How the work is held.
Each of these traces to a defect that shipped, or nearly shipped, in this codebase. None of them is aspirational.
- 01
Availability is not correctness
A caught exception, a non-empty dataframe, a bound port, or an HTTP 200 is not evidence that a result is right. Every one of those has stood in for correctness in a system somebody trusted.
- 02
Refusal is a valid outcome
A system that always returns a number is not more capable than one that declines — it is less honest about the same uncertainty. Refusals carry their real reason and are treated as first-class results.
- 03
An absent measurement says so
A blank beside a filled slot reads as "checked, fine". Unmeasured states are named in those words, because silently defaulting them to zero is how a gap becomes an assertion.
- 04
Name the boundary of every claim
Claims travel with their limits attached rather than in a footnote on another page. Anything that leaves the do-not-say list does so through a recorded adversarial review with sources.
- 05
Keep the negative results
Retreats, refuted claims, and abandoned directions stay in the record beside the wins. Rewriting history to make the current direction look inevitable destroys the only evidence of judgement.
- 06
The decision stays with the person
Alelyon builds models, instruments and assistants. An assistant that acts on a desktop stops and asks before a risky action, and no output is presented as a decision made on someone’s behalf.
Position
On the hardware you already own.
Most capable AI today runs on rented data-centre GPUs and is written for one chip vendor’s software. We are building the other way round: a compute stack on an open standard, models compressed to fit consumer graphics memory, and an assistant that keeps working on the machine in front of you. Where that falls short today, for example in speed against a vendor’s own stack or in the measured quality of a compressed model, we say so.
The tools are bounded on purpose. Every analytical tool Nexus can reach is read-only, and its coding-agent surface can start a child process only after visible user action. Angel’s desktop actions stop and ask before anything risky.
The verification work follows the same instinct. The component that checks a certificate is public and separately installable; the component that computes one is operated. The checker replays scalar and other applicable checks with inputs the recipient supplies and a key pinned out of band. A substrate-sensitive nonzero width may remain explicitly unverified. Replay does not settle whether the data was honest when captured or remove the need to assess the producer; those limits are stated rather than papered over.
Contact operations.
Product, partnership, and technical enquiries reach a mailbox that is read. Nothing about access waits on a person, so this route is for the questions a page cannot answer. No response time is promised until one can be kept.
Contact the lab