Catalytic AITroy, New York

Sovereign models, built machines, and two things you can use today.

Most of our work is one of three things: fine-tuning a model on a corpus that is not allowed to leave a building, building the machine it runs on, or both. Underneath that are two products you can go and use right now without talking to anybody. All of it was trained and benchmarked on the machines in Troy.

Go and use these

Live, and free to try.

LiveSelf-serve

ToxScreen — drug safety AI

Predicts hERG cardiac channel blockade and cytochrome P450 metabolic toxicity from a SMILES string in seconds. Built on the ADMET-AI multi-task architecture, with calibrated uncertainty quantification added — every prediction carries a confidence score rather than a bare label, which matters when you are deciding what to make next.

Built for people screening compound libraries for cardiotoxicity early, before the expensive part of the pipeline. CYP3A4 and CYP2D6 isoform-specific models are in progress.

LiveSelf-serve

Lithic — AI-assisted CNC design

Describe a mechanical part in plain English. Lithic generates CAD geometry as STEP, CAM toolpaths, G-code, and a manufacturing report with estimated cycle time and material utilisation.

Built for small and mid-size shops that need quick turnaround without a full-time CAD/CAM specialist on staff. 3-axis milling, with a tool library you can customise.

Project work

Work that starts with a conversation.

Fine-tuning

Fine-tuning on your own data

Fine-tuning pipeline: curation, formatting, training, evaluation and deployment DATA CURATION PROMPT FORMATTING EVALUATION HARNESS TRAINING · LoRA / QLoRA / FULL DEPLOY · vLLM / GGUF / ONNX ITERATE
Fig. 01 — Fine-tuning pipeline

We handle curation, prompt formatting, training (LoRA, QLoRA or full-weight), and building an eval harness that measures the thing you actually care about rather than the thing that is easy to measure. Models come back as vLLM-compatible checkpoints, GGUF files or ONNX exports.

Our own numbers. We fine-tuned Qwen3-14B on 62,000 examples for Naxi-language translation. It scores 82.3 NTQS, 41 points above GPT-4 on the same eval set. That is our result on our own hardware, and the method behind it is what we bring to other people’s corpora. See the models ↗
Sovereign

Models that stay in your building

For groups who cannot send data to a cloud API at all. We fine-tune and deploy entirely on-premise, on your hardware or ours. Nothing leaves the machine: no telemetry, no third-party API in the loop.

The usual reason is not preference. It is a contract, a regulation or a statute: ITAR and EAR on export-controlled technical data, HIPAA on patient records, CUI under a DFARS clause, attorney-client privilege that a third-party disclosure would waive, or an NDA with a customer whose process data you are not allowed to transmit anywhere. In those cases “we do not train on your data” from a vendor is not the same thing as the data never having left.

Model families: Qwen3, Llama 3/4, Mistral, DeepSeek, Phi-4, Gemma 3. Quantisation to INT8/INT4/FP8 where it helps. RLHF and DPO alignment if the task actually calls for it.

The weights

The checkpoint itself, in a format you can serve without us — safetensors, GGUF or an ONNX export. Yours outright, not licensed back to you.

The pipeline

Curation scripts, prompt templates, training config and the exact package versions. When your corpus grows next year you re-run it yourself.

The eval harness

The one that measures the thing you actually care about, plus the baseline numbers, so a future change can be shown to be an improvement rather than asserted to be one.

The machine

Optional, and often the right answer. Specced, built, burned in and documented — see the reference builds.

What we will tell you honestly. Plenty of tasks do not need a fine-tune at all — retrieval over your documents with a good off-the-shelf model is cheaper, faster to stand up and easier to maintain, and we will say so when that is the case. Fine-tuning earns its keep when you need a house style, a domain vocabulary a general model gets wrong, a structured output format, or a model small enough to run on hardware you can afford to own.
Diagram of a sovereign deployment: corpus, training, evaluation and serving all happen inside your building, evaluation feeds back into the corpus, and nothing crosses the boundary — no egress, no telemetry, no third-party API. YOUR BUILDING CORPUS TRAIN SERVE EVALUATE · RE-RUN IT YOURSELF NO EGRESS · NO TELEMETRY NO THIRD-PARTY API
Fig. 02 — Where the data goes, and does not

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 do.