I build AI agent systems that have to hold up in production: memory, evaluation, permissions, and human review.

• Work: Principal Software Engineer at Expel, building low-latency services over millions of security events and the APIs that put AI into analysts' workflows.
• Before: led an AI group at MITRE (health analytics, federated learning, differential privacy); director at Asymmetrik, where my team built the first secure open-source FHIR server.
• Now also: founder of Intertwine AI, helping teams turn agents into durable colleagues that people can trust.
• Writing: Code & Context, field notes on agent memory, evals, and keeping humans in control.

• dspy-agent-skills: tested DSPy 3.2.x skills for Claude Code and Codex CLI (fundamentals, evaluation, GEPA, BetterTogether, RLM), with validated end-to-end examples. 278★.
• security-verifiers: reinforcement-learning environments for cybersecurity with executable rewards (network-log anomaly detection, config verification), built during Prime Intellect's RL Residency.
• sops-encrypted-envs-mac: turn plaintext .env files into SOPS + age encrypted dotenvs backed by Keychain, with an agent skill and a written threat model.

• The night memory stopped me from doing the wrong thing: an agent, an unauthorized production change, and why approval gates and memory matter.
• Building ClawFleet Academy, an ops control plane for agent fleets
• What local encrypted dotenvs protect, and when to move up: threat-modeling secrets for agent workflows.
• AI memory depends on where the agent lives

I archived two projects in September 2026 after native agent platforms caught up with what they did. That was the right outcome.

• observational-memory: local-first shared memory across Claude Code, Codex, Grok, Hermes and others. Lesson: memory is only useful when it's bounded, inspectable, and tied to where the agent runs. Once the platforms shipped native memory, the right move was to point users there.
• hive-orchestrator (Agent Hive): vendor-agnostic orchestration using Markdown files as shared state, with human-in-the-loop gates on GitHub Actions. Lesson: the durable parts were the human approval gates and the audit trail, not the scheduler.

Both repos stay up, read-only, as reference.

Also a professional bassoonist (bryanyoung.info). Glad to compare notes on agent security, evals, or memory.