Catalytic AITroy, New York

Why the name, why Troy, who is behind it.

The Capital Region has a research university, a semiconductor corridor and a lot of people who need GPUs. What it does not have is anywhere local to get them. That gap is the whole idea.

The name

A catalyst doesn’t do
the reaction for you.

It lowers the activation energy, so a reaction that was impossible at your temperature becomes routine. The reactants were always capable of becoming the product. The barrier was the problem.

That is a fair description of what a 96 GB card does for a student project. The idea was fine. The maths was fine. What stopped it was twenty-four gigabytes of VRAM and a Colab session that dies after twelve hours. Remove that and the same person gets the same result, just without the barrier in the way.

The other half of the metaphor is the hardware itself: a catalytic converter is a honeycomb monolith with a few grams of platinum on it, and it is the single most quietly impressive piece of engineering on an ordinary car. Small, unglamorous, does one job extremely well. We would settle for that.

Section through a three-way catalytic converter: exhaust enters a cordierite honeycomb monolith washcoated with platinum, palladium and rhodium, and leaves converted. EXHAUST IN OUT CORDIERITE MONOLITH WASHCOAT · Pt · Pd · Rh
Fig. 01 — Three-way catalytic converter, section
The idea

Small, local,
and specific.

We are not competing with a hyperscaler on scale, and there is no point pretending otherwise. Two nodes is two nodes.

What two nodes in Troy can do that a region in Virginia cannot: pick up the phone, sit down with you and work out what you actually need, keep your data somewhere you could drive to, and charge a student something a student can pay. For a grad researcher against a thesis deadline, or a shop that has never provisioned a GPU before, that usually matters more than elastic capacity.

We build our own things too — ToxScreen, Lithic, the Naxi translation models — which means when we say a machine suits a job, it is because we have run that job on it.

People

Founder-led, with roles we are hiring into.

KP

Kevin Peters

CEO & founder

Rising senior at RPI, deep in the AI/ML community on campus and connected across student researchers, faculty and clubs. Runs client relationships, campus outreach and product direction.

Active
+

ML engineer

Machine learning

Owns the fine-tuning pipelines, model evaluation, dataset curation and inference optimisation. The person who turns somebody’s data into a model that actually works.

Open
+

Systems engineer

Infrastructure

Owns the hardware: builds, networking, driver stacks, containers, monitoring and on-premise deployments. The person who keeps the nodes up.

Open
+

Community lead

Outreach

Runs the campus presence — posters, club talks, the Discord, faculty outreach. The person who makes sure everyone at RPI who needs a GPU knows we exist.

Open
+

Client engineer

Solutions

Works directly with local businesses and research groups: understands the need, scopes the project, sees it through. Technical enough to spec a server, commercial enough to close.

Open

Interested in one of these roles?

We would rather hear from someone unusual than fill a seat.

What we build on

Whose work this stands on.

Our products apply published methods rather than inventing new ones. It seems worth being explicit about whose work we are standing on.

ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries

Swanson, K., Walther, P., Leitz, J., Mukherjee, S., Wu, J.C., Shivnaraine, R.V., Zou, J.
Bioinformatics, 40(7), btae416, 2024
ToxScreen is built on this multi-task architecture. We add calibrated uncertainty quantification and a single-SMILES web interface. Read it ↗

MoleculeNet: a benchmark for molecular machine learning

Wu, Z., Ramsundar, B., Feinberg, E.N., et al.
Chemical Science, 2018
The Tox21 and SIDER datasets provide the standard splits we benchmark our toxicity models against, alongside internal validation sets. Read it ↗

DeepCAD: a deep generative network for computer-aided design models

Wu, R., Xiao, C., Zheng, C.
ICCV, 2021
Sketch-and-extrude sequence generation is the backbone of Lithic’s geometry pipeline for STEP output. Read it ↗

Text2CAD: generating sequential CAD models from text descriptions

Khan, M.S., Dupont, E., Ali, S.A., et al.
NeurIPS Workshop on ML for Creativity and Design, 2023
Informs our text-to-CAD pipeline. Lithic extends the paradigm with toolpath generation and G-code post-processing. Read it ↗

QLoRA: efficient finetuning of quantized language models

Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.
NeurIPS, 2023
Core methodology for our fine-tuning work, and the reason a single 96 GB card can do what it does. Our PLATINUM-96 and RHODIUM-192 designs are shaped around it. Read it ↗

FlashAttention: fast and memory-efficient exact attention with IO-awareness

Dao, T., Fu, D.Y., Ermon, S., Rudra, A., Ré, C.
NeurIPS, 2022
Used throughout our training pipelines. IO-aware tiling is what makes long-context training fit in 96 GB. Read it ↗

Judging LLM-as-a-judge with MT-Bench and Chatbot Arena

Zheng, L., Chiang, W.-L., Sheng, Y., et al.
NeurIPS Datasets and Benchmarks, 2023
Our NTQS evaluation framework adapts this methodology for domain-specific translation, with human-calibrated scoring. Read it ↗

Megatron-LM: training multi-billion parameter language models using model parallelism

Shoeybi, M., Patwary, M., Puri, R., et al.
arXiv, 2019
Foundational tensor and pipeline parallelism. Informs our multi-GPU configurations with DeepSpeed ZeRO-3. Read it ↗

Come and see the rack.

If you are in the Capital Region and curious, we would honestly rather show you the machines than write another paragraph about them.