Cybersecurity · AI governance · Security automation
I build tools that put security controls, reviewable evidence, and human oversight into AI-assisted engineering. My focus is making automation useful without losing control of what it can access, change, or claim.
Open to opportunities where I can connect security strategy with hands-on engineering and responsible AI adoption.
Connect on LinkedIn · Explore my public work
| Project | What I built | Why it matters |
|---|---|---|
| llm-prompt-guard | TypeScript defenses for prompt injection, untrusted tool content, and potential data exfiltration; published benchmarks and limitations. | AI application security: make trust boundaries explicit and evaluate controls against evidence. |
| bugsweep | Adversarial bug investigation with Hunter, Skeptic, and Referee roles, regression evidence, and review receipts. | Security assurance: challenge findings and require evidence for fixes. |
| nightshift | Recurring AI automation with persistent run ledgers, concurrency locks, approval queues, and scoped change controls. | Operational governance: define authority, retain a trace of work, and coordinate unattended tasks. |
| baton | Evidence-based AI session handoffs that preserve decisions, approval boundaries, uncertainty, and next actions. | Operational continuity: transfer context without treating stale assumptions as current facts. |
| unify-agent-docs | A canonical source for agent instructions, shared invariants, and a pre-commit drift check. | Policy consistency: keep controls aligned across AI development tools and repositories. |
| smokejumper | AI-assisted engineering orchestration with review evidence, deterministic gates, and safety tripwires. | Governed delivery: make review and change boundaries part of the engineering workflow. |
- Layer defenses. Input checks, tool boundaries, output inspection, and human approval address different risks. No single filter establishes that an AI system is safe.
- Make claims testable. Publish limitations alongside results. The prompt-injection benchmark includes missed attacks, not just successful detections.
- Bound automation's authority. Use explicit scope, approval requirements, and recoverable changes rather than assuming an agent should act on every available capability.
- Preserve accountability. Keep evidence of what ran, what changed, what was reviewed, and what remains unknown.
These repositories show my engineering approach to AI security and governance. Each project's documentation distinguishes implemented mechanisms from evaluation results and operational claims.
Product Pilot brings product context into AI-assisted development. Maestro Mobile Testing captures reusable mobile UI-testing workflows.
Portfolio snapshot — checked 2026-11-22 03:49 UTC
| Project | Latest default-branch commit (UTC) | Latest stable release (UTC) |
|---|---|---|
| llm-prompt-guard | 7d35d09 — 2026-11-21 | v2.3.0 — 2026-10-21 |
| bugsweep | 5410ec6 — 2026-10-11 | v0.8.0 — 2026-10-11 |
| nightshift | 892128c — 2026-07-05 | v0.8.0 — 2026-07-05 |
| baton | ebf34de — 2026-09-22 | v2.2.0 — 2026-09-22 |
| unify-agent-docs | 9d5881c — 2026-06-14 | v1.1.0 — 2026-06-14 |
| smokejumper | 4be8041 — 2026-06-10 | smokejumper--v1.0.0 — 2026-06-10 |
Source: public GitHub repository data. Commits may include unreleased work; stable releases exclude prereleases.
Checked daily at 03:47 UTC (GitHub scheduling may be delayed). Only the project-update section is automated. View the updater and run history.


