Launch
Duviloper is live. We're a core AI research and development lab building self-auditing neural systems for high-stakes science — problems where being wrong is expensive, and where a confident-but-wrong answer is worse than no answer at all.
This is the first post on Insights. We'll use it to explain what we're betting on, what we're building, and — just as importantly — what we won't be sharing here.
Here's a failure we keep running into, across domains:
A model is trained on a dataset. It posts a beautiful score on its own test set. It looks brilliant — the kind of number that ends most papers. Then someone collects new data, under slightly different conditions, and the model's skill collapses toward a coin-flip.
Here is the headline, straight from the flagship task: label-free Raman reads drug resistance at ROC 0.89 — against just 0.60 for the sequenced genotype (a 55-gene panel). Raman is reading a phenotypic state of the living cell that a mutation lookup simply misses. The deeper audits — cross-device, cross-site, the confound work — stay gated.
The usual culprit is overfitting — to the wrong thing. Two spectra from different instruments differ in ways that have nothing to do with the sample: a device signature baked into the measurement. An overfit model quietly learns that signature, the easy shortcut, and looks brilliant — right until the instrument changes and the shortcut vanishes. The signal that actually travels is the biology: the part of the spectrum where devices agree. A trustworthy read-out has to track that, not the instrument.
The impressive number was real. It just measured the wrong thing: a shortcut that lived only in the original data, not the underlying science.
This is the core problem Duviloper exists to solve. We don't chase benchmarks. We build systems that interrogate themselves — and tell you, plainly, the accuracy that survives new data, not the one that evaporates the moment a model leaves the lab.
Before any model ships, we stress-test it for the ways AI quietly fools itself. The categories of failure we guard against are not secret — our public site names them openly:
What we publish is that we defend against these. How we test for them, what we find, and how we build around them — that's our edge, and we share it with partners under NDA, not in a blog post.
To prove the method, we picked a hard, real, high-stakes problem.
The problem: deciding whether a bacterium is resistant to an antibiotic — ideally from a single Raman spectrum, a molecular fingerprint read from light scattered off the cell. The goal is to read resistance directly from that signal, sidestepping the one-to-two-day wait that traditional culturing demands.
Antimicrobial resistance is one of the defining health challenges of our era. A wrong answer here is genuinely costly: an ineffective drug given to a resistant infection, or an unnecessary broad-spectrum drug when a targeted one would have worked.
We can say that our method exists, that the model and method are patent pending, and that we describe that we audit it for leakage, confounds, and capacity. How it actually holds up — what it can and can't resolve, and the specifics of why it stays consistent when others don't — is our edge, shared with partners under NDA.
One thing we'll always say plainly: this is research-grade signal, not a cleared clinical diagnostic. Where the clinical evidence isn't there yet, we say so. We will not imply clinical performance we haven't earned.
Most of the AI industry is built on a simple incentive: show the biggest number. We've structured ours differently. A pipeline designed so the model can't quietly cheat — and so we find out fast when a problem isn't solvable yet — is a feature, not a limitation.
When the evidence isn't there, we report it. Not a confident guess. We treat NULL beats noise as a principle: sometimes the honest, valuable answer is "this is not yet knowable from the signal."
That posture isn't modesty. It's the thing we're selling.
We're for:
If you work where being wrong is expensive, tell us the problem. If signal-processing rigor and self-auditing AI can move it, we'll tell you straight — including if we can't.
We read every message. And yes, we're NDA-friendly; we keep your data yours.
Why the cyan underline under the u and i in our name? We code differently, so we spell it differently.
Where is being wrong expensive for you? Bring us a hard problem — we'll give an honest read on feasibility, including if we can't.
Bring us a hard problem →