slaya
Small, fast models that answer yes/no and probability questions so a big AI doesn't have to.
- Status
- Active research
- Year
- 2026
- Role
- Solo
- Area
- Research
In plain English
Many questions we hand to large AI models are really simple decisions, like whether a task is finished or a receipt matches a payment. slaya is my research into small models that answer those questions quickly on the device itself, with a clear yes, no or probability instead of a paragraph.
How it works
- Pick decision points inside my own tools, such as task-done checks, receipt checks and transaction guardrails.
- Collect examples of each decision to build a dataset large enough to train on.
- Use frontier models as teachers that guide and correct the small model, beyond simply generating training data.
- Have the small model answer as a decision tree (a chain of if-then checks) or a calibrated probability (an honest confidence number).
What I did
This is my solo research. I choose the decision points, build the dataset and design the teacher-and-student setup.
What I learned
At this stage, better data matters more than a bigger model. Most of my time goes into collecting clean examples, because a small dataset teaches a model to memorize instead of learn.