Building control-system primitives for supervised AI execution. Software development is the first proving ground — the architecture generalizes to any domain where humans supervise agents in high-consequence workflows.
- Platform Engineering — Cross-platform build systems, CI/CD automation, release engineering, and developer productivity tooling serving 5+ platforms from a single codebase
- AI Execution Infrastructure — Agent orchestration pipelines, RAG systems, MCP servers, execution tracing (OpenTelemetry), reconciliation layers, and human-in-the-loop supervision interfaces
- Engineering Leadership — Defined architecture, standards, and mentoring practices for a 25-engineer organization; drove ~300% release cadence improvement through platform infrastructure and engineering culture
- Open Source — React Native core contributor, ArduPilot autonomous systems contributor, OpenCode ecosystem, and control-surface reliability artifacts
These repos are not independent tools. They are layers of a single control-system architecture.
| Layer | Repo | Role |
|---|---|---|
| Supervision | control-surface-agent | Production reliability control surface: explicit intent, plan, telemetry, reconciliation, operator intervention, and decision artifact in one supervised run. |
| Observability | emerging | End-to-end execution tracing, replay, and longitudinal system introspection (OpenTelemetry) |
| Verification | opencode-devtools | Control-loop gates — type checks, tests, build status as execution constraints |
| Execution | OpenCode | Agent runtime (contributor — RLM REPL sandbox fix, Learn plugin) |
| Learned Policy | opencode-learn | Feedback-driven adaptive controller — agents learn adaptive control hints, not ground truth |
| Epistemic Safety | agent-memory-failure-demo | Controlled experiment proving lossy memory degrades agent correctness |
| Thesis | Control Systems for Intelligent Software | Architecture essay series — execution, verification, observability, control surfaces, supervision |
Human Intent ──> Orchestration ──> Agent Execution ──> Verification ──> Human Supervision
│ │ │
task routing tool calling trace trees
context mgmt parallel exec replayable
priority queue checkpoints sessions
The same architecture that governs robotics and industrial automation (I'm an ArduPilot/MAVLink contributor) applies to AI agent systems: machine telemetry, reconciliation, operator interfaces, and human-in-the-loop control.
Verification is bounded. Execution is open-ended. Reliability emerges through supervision, telemetry, and correction across repeated runs.
Software development is the current implementation context — it's observable, replayable, rich in intermediate artifacts, and a strong sandbox for control-loop design. But the primitives generalize:
| Domain | Agent executes... | Human supervises... |
|---|---|---|
| Software | code edits, test runs | merge decisions, architecture |
| Research | evidence gathering, synthesis | source evaluation, conclusions |
| Operations | workflow execution | exception handling, escalation |
| Analytics | data traversal, pattern detection | decision validation |
The architecture is always: intent → orchestration → execution → verification → supervision.
Safety in staging is a checkpoint. Reliability in production is a continuous control problem.
The anchoring constraint across every domain: never destroy raw evidence. Summarize for navigation, not for storage.
| Platform Engineering | CI/CD automation, release engineering, cross-platform build systems, deployment infrastructure, OTA updates, developer productivity tooling |
| AI Infrastructure | Agent orchestration, RAG pipelines, MCP servers, execution tracing (OpenTelemetry), LLM applications, prompt engineering |
| Systems & Cloud | AWS serverless, microservices, DDD, multi-tenant SaaS, Docker, Kubernetes, infrastructure as code |
| Languages & Runtime | TypeScript, Python, C#/.NET, Node.js, React/React Native, GraphQL |
- React Native — Contributor to the core framework
- ArduPilot — Contributor to autonomous vehicle/drone firmware and documentation
- OpenCode — Submitted fix for critical RLM REPL sandbox crash (experimental PR), built mobile supervision client, created knowledge persistence plugin
Series: Control Systems for Intelligent Software — mapping robotics and industrial automation architecture onto AI agent systems.
- Bash Is All You Need — Until It Isn't — Execution is solved. Context curation, orchestration, verification, and supervision are the real problems.
- Why the Smart Model Reviewer Pattern Is Backwards — Strong models should generate. Cheap models should verify. The industry has it backwards.
- AI Agents Are Control Systems — The ArduPilot/MAVLink architecture is a proven blueprint for AI agent supervision.
- Observability for AI Agents — Traces and replay are how you tell whether a system remains reliable as it accumulates runs, interventions, and drift.
- Why AI Needs Control Surfaces, Not Just Chat — Chat is fine for intent. Reliability needs control surfaces that make system behavior legible enough to supervise across runs.
- control-surface-agent — Production reliability infrastructure for agentic systems. A bounded decision workflow with explicit intent, explicit planning, telemetry, reconciliation, operator intervention, and a decision artifact.
- agent-memory-failure-demo — Controlled experiment showing how lossy memory degrades agent correctness.