Research
Applied work on small models, local hardware and agents that act.
Three questions we are paid to answer, and one we are not.
Local inference
How much teaching fits in an open model
Lesson plans, notice drafting and translation on leading open-source weights, hosted on hardware we operate. We measure usefulness per watt, not benchmark scores.
Agents that act
Screenshot plus hierarchy, then a tool call
A vision model sees a frame and a compacted UI tree each turn, then taps, types or scrolls. The interesting failures are perceptual, not linguistic.
Integrity
Detecting copied homework from a photo
A similarity pipeline over uploaded images that survives crops, filters and rotations, and reports a signal for a teacher to judge — never a verdict.
Not our work
Training frontier models
We buy that capability by the credit when a student genuinely needs it, and spend our own hardware on the work that must stay private.
Our hosting principles, in plain language
Hosting
Local
Model approach
Open source
Workloads
Text + vision
Privacy
By design