We put a model on a machine you own, then we leave.
Some work cannot go to somebody else’s API: export-controlled drawings, patient records, a corpus an NDA keeps in one building. That is the work we take, and most of it is boring. Build the machine, train on the data, measure the result, hand it over.
The name comes from the machine, not the marketing.
That describes most of the job. The data exists, the question is already well posed, and the team knows what a good answer would look like. What has been missing is a way to get one without sending the corpus somewhere it is not allowed to go.
Remove that constraint and the result is reachable. One rack upstate, doing one class of job properly. The firm is not trying to be large.
Small, specific,
and in one building.
We cannot compete with a hyperscaler on scale, and there is no version of this where we could. We sell the other thing: a model that belongs to one organisation, sits on one machine, and keeps working whether or not a vendor stays in business or holds its price.
The corpus stays in your building, and the weights come to you in a format you can serve without us. Nothing in the request path is ours. When the machine needs repair, the person who built it handles it. For a defence supplier under ITAR, a hospital working under HIPAA, or a shop that has never provisioned a GPU, that counts for more than elastic capacity.
We build our own things too — ToxScreen, Lithic, the Naxi translation models — so when we say a machine suits a job, it is because we ran that job on it. When we say a fine-tune will not fix your problem, we have usually tried the same thing ourselves.
Small on purpose. You deal directly with the people doing the work.
We are a small firm and have no interest in appearing larger. On every project you speak with the people who will do the work; there is no account manager relaying messages. A fixed-scope fine-tune is $7,500 and the evaluation before it is $2,500, both quoted before anything starts.
We also turn work down. A fine-tune is the wrong answer more often than people expect: retrieval or a smaller change is cheaper and quicker to stand up, and saying so in week one costs you nothing.
Roles we are hiring into
Scoping and model evaluation
The first job is deciding whether a fine-tune is the right answer at all. Often it is not. A retrieval system or a simpler change gets there for less money. Better to say that in week one than to find it halfway through a training run.
ML engineer
Owns the training pipelines, the eval harness, dataset curation and inference speed. Turns somebody’s data into a model that holds up on data it has not seen.
Systems engineer
Owns the hardware: builds, networking, driver stacks, containers, monitoring, air-gapped installs. Shows up on site with a torque driver and leaves a machine that survives an audit.
Deployment engineer
Stands a finished model up on the client’s machine: containers, auth, logging, the handover session. Usually the only person in the building who can read a stack trace.
Client engineer
Works with the client directly: frames the need, scopes the project, and sees it through. Technical enough to spec a server and to review a training run, and expected to do both.
Interested in one of these roles?
We would rather meet someone unusual than fill a seat.
What we borrowed, and from whom.
Our products apply published methods. We did not invent the architecture under ToxScreen or the tuning method behind the Naxi run. The people who did deserve their names on it.
ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries
MoleculeNet: a benchmark for molecular machine learning
DeepCAD: a deep generative network for computer-aided design models
Text2CAD: generating sequential CAD models from text descriptions
QLoRA: efficient finetuning of quantized language models
FlashAttention: fast and memory-efficient exact attention with IO-awareness
Judging LLM-as-a-judge with MT-Bench and Chatbot Arena
Megatron-LM: training multi-billion parameter language models using model parallelism
Come and see a build log.
If a conversation would settle it faster than a form, say so and we will find half an hour. We would rather show you a build log than write another paragraph about one.