Building AI systems, agents, and Physical AI - with a focus on local-first, practical intelligence.
Lately I've been working on authority and safety architecture for autonomous agents: how to let an agent act with real capability without giving it, or an attacker who compromises it, unrestricted power.
- Sentinel - an authority-attenuating immune architecture for AI agents. Detects and contains compromised agents without giving defenders unrestricted authority.
- Capgate - runtime authority boundary for autonomous AI agents. The agent proposes, the harness authorizes, the executor acts. Deny-by-default policy, non-transitive delegation, tamper-evident audit.
- Agent-Skills-Compendium - an open compendium for discovering, defining, composing, and governing reusable capabilities for AI agents, from digital intelligence to multi-agent systems and Physical AI.
- QUANTUM-CUDA - numerical verification and the patch behind NVIDIA/cuda-quantum#5459, documenting exp_pauli, givens_rotation, and fermionic_swap.
- NVIDIA/physical-ai-data-factory#22 - fixed a
NameErrorcrash in the Video Data Augmentation pre-submit guard's model-cache check. - NVIDIA/cuda-quantum#5459 - documented
exp_pauli,givens_rotation, andfermionic_swap.
- Agent authority and permission boundaries
- Multi-agent governance and skill composition
- Physical AI
- Local-first, practical intelligence over hype
- Website: https://www.jersonboydmilan.com