The research that
powers research.
We study how machines can think better, use less energy, and serve people without surveilling them.
The environmental cost of cloud AI.
Every cloud AI query travels to a data center, consumes GPU power, generates heat, and requires cooling. Millions of queries per day means millions of kilowatt-hours burned for text that could be generated on the device in your hands.
Local AI eliminates the round-trip. Your Mac's neural engine does the work. No data center. No transmission loss. No cooling infrastructure. The greenest compute is the compute that doesn't travel.
Smaller models. Smarter context. Less waste.
The industry trend is bigger models, more parameters, more compute. We're going the other direction — models that are small enough to run on consumer hardware but smart enough to be useful, because they understand YOUR context instead of trying to know everything about everyone.
A 4-billion parameter model that knows your field outperforms a 70-billion parameter model that doesn't. Context beats scale. Every time.
What happens when AI can't see your data?
Most AI research assumes access to user data — for training, for improvement, for alignment. We ask: what if the AI never sees aggregate data? What if improvement happens locally, per-user, without any information leaving the device?
This constraint forces better architecture. You can't rely on scale when you can't collect data. You have to be smarter about how the model adapts, how context is maintained, and how personalization works without surveillance.
What we're working on.
How do you make a local model that improves over time without sending data anywhere? How do you build personalization that's portable between devices but readable only by the user? How do you reduce the energy footprint of AI by 100x while maintaining quality?
These are the questions that drive our work. When we have answers worth sharing, we'll share them here.
Interested in what we're building?
We're a small team solving hard problems. If that's your kind of work, reach out.
Get in touch