Why stuffing a million-token window causes attention rot and quadratic billing, how SentencePiece vs. BPE tokenizers silently break budgets, and how stratified three-tier memory graphs keep agents sharp.
Robotics, embedded Linux, and applied AI systems.
Personal engineering workbench of Shawn Maybush. Focusing on closed-form kinematics, hardware abstraction layers, containerized toolchains, and distributed autonomous agents.
Recent Lab Notes
Why natural language begging rots in production, and how bounded workspaces, deterministic file primitives, and actionable error contracts make autonomous agents reliable.
Using dual quaternions and JAX-accelerated polynomial regression to solve inverse kinematics for robotic manipulators with low-overhead edge inference.
Creating an isolated ROS2 development and testing platform with C++ I2C drivers, i2c-stub emulation, and Docker container passthrough on Raspberry Pi.
Automating code reviews and documentation updates within a GitLab CI/CD workflow using Google Cloud Functions and Vertex AI.