Writing
Notes on model training, agents, observability and agent architecture.
Supercharging a Service with the Provisional Ability of MCPs
Better tooling happens when MCPs balance the cognitive load for client agents.
A Grounded Voice Copilot for 3D Vehicle Inspection
Mobisim turns a raw GLB vehicle into something you can talk to using structural DAG extraction, a reviewed semantic overlay, real-time WebRTC, and deterministic execution.
Fighting Specification Gaming with Verifiable Rewards for Evaluations
Designing an RLVR-style verifier and scalar reward function for agent evaluation workflows, with judge metrics, render-validity gating, and reward diagnostics for stability, win rate, and failure modes.
Building Models for Cellular Automata Prediction
Research notes on modeling stochastic cellular automata with entropy-based patching and transformer architectures for next-state prediction.