Quaternion Studios — Applied AI & Interactive Engineering
Applied AI, embodied autonomy, and real-time 3D — researched, engineered, and shipped to production for teams working on genuinely hard problems.
Ethos
A quaternion rotates through a dimension you can't see — the axis between the real and the imaginary. That is where the interesting engineering lives: not in the model and not in the product, but in the narrow space between them where the two have to actually agree.
It is a small space, and not many people work in it. Research groups build things that are true but not usable. Product teams build things that are usable but not true. We sit in the middle and make the two meet, which is slower to explain and considerably faster to ship.
How we workWhat we do
Six things we are genuinely good at. If your problem is not one of them, we will tell you that in the first conversation rather than the third invoice.
Retrieval, semantic search, and multi-agent systems paired with real domain solvers — models that reason against ground truth instead of around it.
Vision-language-action policies, imitation learning, and the inference plumbing that gets a 3B-parameter policy running inside a 20 ms control loop.
Three.js, WebGPU, GLSL, WASM — configurators, BIM tools, and visualizations that render thousands of elements without dropping a frame.
Large, messy datasets turned into fast, interactive, GPU-accelerated views that stay responsive at scale rather than at demo size.
Production backends under load — transaction engines, knowledge graphs, distributed processing, cloud automation, and the observability around them.
Constraint-aware CAD editing, BIM/Revit pipelines, and automation that turns hours of manual drafting into minutes of review.
Practice areas
Four bodies of work we keep returning to, and the specific problems inside each one that we have had to solve more than once.
Policies that move real hardware. We work across the whole stack — from what the robot sees to what the actuators do 20 milliseconds later.
Most agent failures are not model failures. They are context, contract, and evaluation failures. We design for those first.
Interfaces that hold tens of thousands of elements and still feel immediate. The frame budget is the specification.
The unglamorous half that decides whether any of the above survives its first real week in production.
Selected work
Journal
Long-form engineering writing on vision-language-action models, robot autonomy, multi-agent systems, and the rendering work underneath experiences like this one.
Embodied AI · 9 min read
VLAs collapsed perception, language, and control into one network in about three years. Here is the lineage, the three axes every design choice falls on, and the decisions that actually determine whether yours works on hardware.
21 Jul 2026Embodied AI · 6 min read
A 3B-parameter policy cannot run at 50 Hz, and yet dexterous manipulation needs 50 Hz. Action chunking is how that contradiction gets resolved — and temporal ensembling is how you pay for it without the seams showing.
7 Jul 2026Embodied AI · 7 min read
Denoising an action sequence is an unreasonably good way to model multimodal behaviour. It is also a sampling loop inside a control loop, which is where the engineering starts.
23 Jun 2026How we work
Engagements differ in subject matter far more than they differ in shape.
Before a line of code, we find out what is already known. Most 'novel' problems have three papers and a reference implementation behind them, and the fastest path is knowing which one is load-bearing.
A narrow, honest prototype against your real data — not a demo dataset. It either shows the approach works or it kills it in days instead of quarters. Both outcomes are worth paying for.
Evaluation harnesses, failure modes, latency budgets, cost ceilings, observability. This is where a promising notebook becomes something you can put in front of users.
Documented, instrumented, and legible to your engineers. We would rather your team owns it in six months than that you keep paying us to be the only ones who understand it.
Elsewhere on this site
The Field is the same studio, laid out as a surveyed plane instead of a page. You load standing in an ink-grass hero, rise out of it, and land on a map: one origin and six destinations, each exactly the same distance out, with a real-time fractal lab and the case studies orbiting in between.
Everything out there is rendered live. Nothing out there is decoration.
Enter the FieldLet's build
Tell us what you're solving — applied AI, embodied autonomy, real-time 3D, or the systems underneath. We read every message, and we'll tell you honestly if it isn't ours.