I build local-first, auditable systems for AI agents and evaluation workflows. The work focuses on observable agent runs, deterministic verification, bounded authority, reproducible proof, and handling external content as untrusted data.

DIF Defense: behavioral prompt-injection detection for bounded agent workflows, using observable traces and explicit clean, warn, or compromised verdicts.

Terminus AEGIS Protocol: safe web traversal and observability with frozen user intent, hardened ingress, receipted evaluation, Trust Reports, and replay.

Local Eval Foundry: deterministic, provenance-bound evaluation infrastructure with reproducible benchmark packs and auditable run receipts.

Agent Flight Recorder: local-first observability for tool-using agents, covering model calls, tools, state, checkpoints, replay evidence, and regression cases.

cindermote: a disposable microVM runtime for untrusted agent material.

My GitHub includes the source repositories and technical context for this work.