Robotics / AI / Systems
Selected work in open robotics, physical AI, healthcare platforms, and retrieval systems.
01 / Selected work
The projects combine software and hardware, often with uncertain data.
A framework for moving from an idea to working hardware without rebuilding the same infrastructure every time.
It defines clear interfaces across software, firmware, and hardware so each layer can change without rebuilding the rest.
Read the project storyFulmetry gives your coding agent the project structure and commands to write the TypeScript, place and route the PCB, then build and test it.
You review the schematic, PCB, and 3D assembly in a local workspace, steer the next change, and prepare the finished board for manufacturing.
Read the project storyAn autonomous agricultural platform combining spatial reconstruction, visual pest detection, and robot arm control.
Gaussian Splatting informed navigation, CNNs handled pest detection, and forward/inverse kinematics drove precise manipulation.
A retrieval approach that uses Monte Carlo exploration to search a corpus rather than depending entirely on vector similarity.
Early experiments reached roughly 80% recall on the target task, compared with under 20% for the embedding baseline. The experiment examines the search behaviour behind that gap.
Clinical platforms, remote patient monitoring, real-time communication, claims workflows, and document systems assisted by AI.
Work across startups and scaled healthcare companies focused on reliability beyond the happy path.
Other projects
Tools and experiments you can inspect or run.
A TypeScript memory engine combining BM25, embeddings, recency, and feedback with adaptive weighting for each query and an L1/L2/L3 memory hierarchy.
A working assistant for small Indian businesses: transaction capture, cash flow forecasts, payroll, collections, and proactive alerts through a familiar interface.
02 / Working questions
These questions guide what I build next.
The useful kind of boring: predictable, inspectable, and reusable.
Search may need strategies, memory, and backtracking alongside better embeddings.
Documentation and interfaces matter as much as a downloadable CAD file.
03 / Field notes
Causal attention, GQA, FlashAttention and RoPE answer different questions. Seven categories show which part each one changes.
A loop-engineered robotics harness should route experiments through durable state, independent verification, simulation, replay and supervised hardware.
Mark Gaussian splats with object identity, retrieve the relevant scene evidence, and let geometry handle the drive.
A short signal