The model runs on a machine in your building.
We do two things. We fine-tune a model on a corpus that is not allowed to leave a building, and we build the server it runs on. Often both, because the two problems come up together. Evaluation is $2,500, a fine-tune $7,500, a model with its machine from $25,000. Two small products at the bottom of the page are free to open today.
On evidence: our Naxi fine-tune was measured on a 446-row held-out gate with paired bootstrap confidence intervals clustered by reference, and reported on two independent metrics — 45.78 chrF++ and 69.07 on NTQS, a five-part translation rubric we wrote. The weights, the rubric and the scoring script are on Hugging Face, so the number can be reproduced rather than taken on trust.
The reason is usually
a contract.
We work with groups who cannot send data to a cloud API at all. Both the training run and the serving happen on-premise, on your hardware or in our rack. No telemetry. No third-party API anywhere in the loop.
The clause is usually a statute or a contract. ITAR and EAR on export-controlled technical data. HIPAA on patient records. CUI under a DFARS clause. Attorney-client privilege, which a third-party disclosure waives. Or an NDA covering a customer’s process data. When one of those applies, a vendor promising not to train on your data is a weaker thing than data that never left the room.
What that produces for an assessment. Each engagement closes with a written record: model provenance, the evaluation results with the methodology and confidence intervals stated, access and retention controls, and the reproducible build details for the hardware. Those artifacts are framework-agnostic. They are designed to be mapped to whichever framework applies to you — the NIST AI Risk Management Framework, a state AI statute, a sector rule, a procurement questionnaire or your own internal audit — and we do not represent any of it as legal compliance. The decision about what is sufficient stays with your counsel and your auditors.
Model families: Qwen3, Llama 3/4, Mistral, DeepSeek, Phi-4, Gemma 3. Quantised to INT8, INT4 or FP8 where that helps. DPO or RLHF where alignment work is genuinely part of the task.
Two engagements. Most projects need both.
Fine-tuning on a corpus that cannot leave your building
Curation, prompt formatting, training (LoRA, QLoRA or full-weight), and an eval harness that measures the thing you care about. You get the model back as a vLLM-compatible checkpoint, a GGUF file or an ONNX export.
Data handling is agreed before any of that starts. Where the corpus sits, who can see it, which machine holds it, what happens to it when the project ends. Most clients keep training on their own hardware and we work on it under supervision. Where that is impractical we run it on ours, isolated, and hand over the run logs.
Running on your machine
The model serves inside your network, on hardware you control, on an inference stack you can read. vLLM for throughput. TensorRT-LLM where it earns its keep. llama.cpp and GGUF on small machines.
How it lands in practice.
- Your hardware, your team runs it. We build the model and the stack and document both. You keep the manual.
- Your hardware, we stand it up. On-site install, burn-in, wired into whatever auth and logging you already run, then support if you want it.
- Air-gapped. A machine that has never touched a network, the model already on it, and an audit document listing what is in it and how it was built.
In every case the weights are yours outright. No licence back to us. No endpoint of ours in the request path.
The weights
The checkpoint itself, in a format you can serve without us: safetensors, GGUF or an ONNX export. Yours outright.
The pipeline
Curation scripts, prompt templates, the training config and the exact package versions. When the corpus grows next year you re-run it yourself.
The eval harness
The one measuring the thing you care about, plus the baseline numbers. A later change can then be shown to be an improvement. An auditor or an IRB asks for exactly this, and arguing a change is better is how those conversations go wrong.
The machine
Optional, and often the right answer. Specced, built, burned in and documented down to part numbers. See the reference builds.
The machine it
runs on.
A model nobody can operate is not deployed. We spec the server, source it, build it, burn it in, install the stack, and hand it over with the model already serving, or we work alongside your team until they can operate it. Which of those you want is settled at scoping.
Reference builds →Sized to the job
A 7B model at 4-bit does not need a 96 GB card. We work back from the model and the query load to the smallest machine that holds it with headroom, and we tell you when that is a $4,000 box instead of a $15,000 one.
Documented for an audit
Parts list with part numbers, thermal readings under sustained load, the software bill of materials, and a build guide somebody else could follow. That paperwork is the difference between a server and a black box.
Stood up and left working
OS, drivers, CUDA, containers, monitoring and the inference server, installed and checked before it ships. It arrives working. A box with a wiki link is not the same thing.
Support if you want it
Optional and priced separately. Hardware fails eventually. When it does, you are talking to whoever put the machine together.
Live, and free to try.
Two narrow models we built for our own work and left available publicly. Each covers one corpus and reports a confidence score with every output. No account needed.
ToxScreen — drug safety AI
Predicts hERG cardiac channel blockade and cytochrome P450 metabolic toxicity from a SMILES string in a few seconds. Built on the ADMET-AI multi-task architecture, with calibrated uncertainty on top: every prediction carries a confidence score.
For people screening compound libraries for cardiotoxicity early, before the expensive part of the pipeline. CYP3A4 and CYP2D6 isoform-specific models are in progress.
Lithic — AI-assisted CNC design
Describe a mechanical part in plain English. Lithic returns CAD geometry as STEP, CAM toolpaths, G-code and a manufacturing report with estimated cycle time and material use.
For small and mid-size shops that need a quick turnaround without a full-time CAD/CAM specialist on staff. 3-axis milling, with a tool library you can customise.
Not sure which of these you need?
That is a normal place to start. Describe the problem in your own words and we will tell you which of these fits. Or that none of them does.