The problem, not the model
A model is a part, not the point. What matters is whether the thing solves the problem, not how it scores on a benchmark nobody outside this field has heard of.
We're two students building practical AI and we're the first users of everything we make.
Plenty of people build models. Fewer make them usable. We spend our time on the second part: gluing speech, vision, language and search into things you can open and use without reading a manual first.
What we believe
A model is a part, not the point. What matters is whether the thing solves the problem, not how it scores on a benchmark nobody outside this field has heard of.
Speech, vision, language and search are the same four pieces every time. What changes is the arrangement. That's why a new tool takes us weeks instead of starting over.
An answer should point at where it came from. If you can't check it you're just trusting it, and there is enough of that going around already.
Open models and systems we control. Nothing here depends on one company staying in a good mood about its pricing.
The people
Works on making sense of multimedia: speech, frames, and everything people leave unsaid in between. Also likes cookies.
Works on retrieval, making hours of video as searchable as a text file.
The bigger picture
NNL isn't one product. NINA turns recordings into notes you can move around in. IRIS finds the exact moment something was said. LUNA turns a pile of scattered files into a single document. ARIA is a chat that keeps the thread. Four different problems, and underneath, mostly the same machinery.
Each one teaches us something the next one uses. That's the whole plan, and it's deliberately a small one: build the next useful thing, learn from it, build the one after that.
Let's build
Tell us what you're working on and we'll try to help. No fees, no catch. We're two people who like building things and we're curious what you'd point these at.