Thinking in public
Notes on AI, robotics, hardware, and the intersection of digital intelligence with physical systems. Written as we build, not after.
RSS & updates coming soon
Follow the GitHub for now — research notes often ship alongside code.
Published
4 postsWhy closed-loop intelligence matters more than model size
A working 7B model that receives sensor feedback and updates its beliefs in real time is more useful than a 70B model that acts on stale priors. Here's why the loop is the real innovation.
Designing robots to be taught, not programmed
Traditional industrial robots are programmed via teach pendants and waypoints. AI-controllable robots need to be designed differently — from the actuator selection up. Notes from building the Adeept arm integration.
3D printing as a scientific instrument
When your lab has a 3D printer, every experiment can have a custom fixture in 4 hours. The design-to-test cycle collapses in ways that change how you think about rapid science.
Building AI agents for materials discovery
Notes on running autonomous inverse design campaigns for MXene and perovskite compositions — including the dataset bias problem, balanced sampling, and what it means for an agent to 'propose an experiment'.
Coming up
2 draftsOn interpretability in physical AI systems
When an AI drives a robot arm, you need to know not just what it's doing but why — and how to override it safely. This is harder than interpretability for pure software systems.
Rapid prototyping as a research methodology
A defense of building things fast and breaking them. Why physical prototyping speed matters for AI research, and how to set up a lab workflow that makes it the default.