Independent AI research.
An independent research lab working on language models, simulation platforms, and the developer tools that surround them. Scope is deliberately narrow.
Quiet research. Working tools.
Venode is a research lab. We design language models, simulation platforms, and the tools that surround them.
How the lab works.
01 · principlesResearch before product.
Every model we ship is written up first. The journal documents the design choices, the evals, and the failure modes. Reasoning ships before the product does.
Tools that stay out of the way.
We build for the desk, not the demo. Plain reading on every surface, no house style imposed on the user, no padded answers, no enthusiasm on someone else's behalf.
What the lab builds.
02 · productsHugO
The venode language model. A cloud service shipped in four variants, each tuned for a specific class of work. One API, one persona, the right variant for the job.
- Four purpose-trained variants, one base model
- 32K context across every variant
- Editorial restraint by default
Plenum
Fire engineering simulation. CFD-driven smoke modelling, evacuation pathfinding, and compliance checks under AS 1668.1 and NCC Spec 17. In design, with the first real scene due later this year.
- CFD smoke flow as buoyant turbulence
- Egress pathfinding on the floorplan
- AS 1668.1 and NCC Spec 17 verdicts
Xact
Estimation software for the construction and trades services. Itemised cost build-up from quantities, materials, and labour, with markup applied per line, per trade, or per project. In design.
- Itemised estimates from quantities and labour
- Markup configured per line, per trade, or per project
- Take-offs from drawings, with rate libraries
- Quote, invoice, and contract exports
Custom builds · on your terms
Built for you. Not trained on you.
If you don’t want a shared product, we build one for you. On-site, hosted, or hybrid. We design and train on your corpus, ship the result to your stack, and stay out of your data. No telemetry into a shared pool. No advertisers. No resale, ever.
-
Custom build, your product.
We build to the shape of your work. Trained on your corpus where it’s a model, wired to your stack where it’s a platform, named whatever you want it named.
-
On-site, hosted, or both.
Run it on your hardware behind your firewall, or on ours, or split the difference. The choice stays yours.
-
Your data does not train anything else.
We do not pool customer data into a shared training set. Your corpus trains your model and only your model.
-
No advertisers. No resale.
We do not sell data, sell access to data, or hand it to a third party. Not for revenue, not for partnerships, not at all.
Recent notes from the desk.
03 · journal
N-014
N-013
N-012
N-011
On reading before writing · Hugo's editorial prior.
Why the assistant should hold the document before it answers, and what falls out when it does. The case for long context as the default, not a tier upgrade.
Pareto-tuned variants, one base.
A short write-up of the variant fine-tuning approach: one base, four variants, one eval set per variant, one shared persona check. What we tuned for and what we did not.
Plenum: solving smoke flow as buoyant turbulence.
Sketch of the low Mach number Navier-Stokes setup we use inside Plenum, including the radiation closure we settled on after three rejected attempts.
What a one-desk lab can ship in a year.
A quiet year-end stock take: what reached private preview, what failed silently, what got rebuilt twice. Numbers, not vibes.