MundiLABS
Computer vision · Language models · Robotics
Mundi Labs is an independent AI practice. We take perception, language and robotics systems from a result that holds on a benchmark to one that holds in the field — and we publish what we learn along the way.
The problem
The benchmark is clean, balanced and finished. The world is none of those things. Lighting changes, the camera moves, the user phrases it differently, the gripper slips, the distribution drifts, and the number that looked settled in the paper quietly stops holding.
Most teams find this out in production. They have a result they trust, a demo that impressed the room, and no instrumentation between that and the thing their customers actually touch. The fix is rarely a bigger model. It is measurement that reflects the deployed setting, and the engineering discipline that measurement makes possible.
We work on both sides of that gap, because we do not think they are separable.
Vision, language and embodiment stopped being separate disciplines some time ago. We work across all three, and most engagements touch at least two.
Perception that holds under conditions the training set never contained — new sensors, new sites, new lighting, new failure modes. Detection, segmentation, depth, tracking, reconstruction and the evaluation to prove any of it moved.
Retrieval, agents and tool use taken from a prototype that demos well to a service that holds under real load, real users and a real cost ceiling — with evals that tell you the truth about whether a change helped.
Policies that survive contact with hardware. Manipulation and navigation, the simulation-to- reality gap, and the unglamorous instrumentation that tells you why the robot failed at 3 a.m. rather than that it did.
The connective tissue, and the reason to hire one team instead of two: we take a method from the literature — sometimes our own — and turn it into something you can deploy, maintain and explain.
Every engagement runs the same five stages in the same order, because skipping to the fix is how teams end up optimising something they never defined.
We agree what “working” means in your terms — not accuracy in the abstract, but the specific outcomes and failures that carry cost for your business — and what evidence would settle the question either way.
We put measurement in place before touching a model: an evaluation set drawn from your deployed distribution rather than a public benchmark, graded against human judgement, and a harness your team can run on demand.
We baseline the current system, sort the failures by what they actually cost you, and tell you plainly which are worth fixing and which you should accept. Some engagements stop here, and that is a fine outcome.
We work the list in priority order — architecture, data, training, retrieval, control — with every change gated on the harness, so improvement is demonstrated rather than asserted.
You keep the harness, the datasets, the documentation and the working knowledge. The test of a good engagement is that your team can run the next one without us.
Consulting firms that only consume research drift, slowly, into applying methods they no longer understand the limits of. Writing for peer review is how we keep that from happening — it forces us to state what we actually claim and to be wrong in public when we are.
It is also what you are buying. When a method comes out this month, we can tell you whether it applies to your problem, because we read it properly and often reproduced it.
Where an engagement produces something genuinely new, we will ask to publish it with you. Where it touches anything sensitive, we will not.
Fixed scope and fixed fee wherever the work allows it. We would rather tell you the engagement is unnecessary than sell you a longer one.
A short, independent read on a system you have already built. We instrument it, baseline it against your real distribution, and hand you a written verdict with a ranked list of what to do next.
We embed with your team and ship alongside them — writing code, reviewing designs, and raising the floor of the whole system while it goes to production.
A retained line to call when the decision is expensive and reversible only at cost: architecture review, method selection, vendor diligence, or a launch that needs a second opinion.
We reply to every serious enquiry within two working days, usually with questions before a proposal. First conversations are free and frequently end with us saying you do not need us yet.