Why this

How small can a robot’s mind get—and still do something useful?

I’m Shubham. I started shubh.ai as a one-person AI venture: a place to build practical things with existing models and share what I learn. That idea has become more specific. I’m working on small, specialized models that run on a robot’s own hardware.

A robot has to act in a world that keeps moving. A balancing controller needs to respond before the machine falls. A grasp depends on the camera, the control loop, the motors, and the object itself. A capable model is only one part of that system.

I’m interested in the space between a model and a motor: turning what a model knows into timely, reliable action. That means working on distillation, efficient runtimes, control, and the hardware they actually run on.

Small, local, embodied.

Small means finding the simplest model that can do a task well. Local means making it work on the board attached to the robot, within its memory, power, and timing limits. Embodied means testing it on a real machine. Simulation helps me learn; hardware shows me what I missed.

These are working constraints, not results I’ve already achieved. I want to measure the whole system: task success, response time, power draw, and cost. A fast model is useful only if the robot can use it.

Build, test, share.

I work independently to keep the loop between an idea and an experiment short. I don’t need to train another foundation model to begin. There is plenty to do with the models and tools already available.

I’ll share experiments, small open-source models and tools, and notes that make the work easier to understand—including what didn’t work. The aim is simple: make useful robot behavior possible with less, and leave enough detail for someone else to build on it.