Cybersecurity Engineering @ George Mason University · B.S. Dec 2028 · Minor in Data Science AWS Certified Security - Specialty · studying AWS DevOps Engineer - Professional
I build systems that assume something will go wrong, and write the tests that prove what happens when it does. Most of what is here is infrastructure for AI coding agents and cloud security work on AWS. Lately I spend my competition time on the other side of that: red-teaming AI agents.
| Project | What it is | Scale |
|---|---|---|
| provalume · PyPI | Evidence-based memory for AI coding agents. Only trusts what was proved: a check that passed, a reviewer who was not the author, a commit that actually landed. Append-only hash-chained journal, deterministic rebuild, five-rung trust ladder. | 26k LOC · 566 tests (161 security) · 26-threat model · 18 ADRs |
| orkestra · PyPI | Runtime that coordinates many coding agents (Claude Code, Codex CLI, Antigravity) on one repo. Each task in its own Git worktree, your tests are the referee, and nothing lands until a different agent has reviewed it. | 16k LOC · 557 tests · 17 releases · 15-threat model |
| Secure Data Hub | Multi-account AWS architecture for client PII. Cross-account STS, KMS item-level encryption, tenant identity taken from a validated JWT claim rather than the caller. | 9-account design · 4 least-privilege IAM policies · 31-slide architecture deck |
| tableau-public-authoring-mcp | MCP server that compiles a declarative spec into Tableau workbooks and validates them against the real desktop app. | 17k LOC · 419 tests, more test code than source |
| local-ai-orchestrator | Seven-agent local LLM pipeline with a deterministic validation layer that overrides model self-assessment. Runs offline, no paid APIs. | 11k LOC · 215 tests |
| AGA_Datathon_2026 | 🥇 1st place, National AGA Datathon. I built the public site and the methodology appendix that made federal spending and audit data legible to ordinary citizens. | 8 pages · 29 appendix documents |
Cloud security on AWS. IAM across trust boundaries, cross-account sts:AssumeRole, KMS
customer-managed keys, Cognito and OIDC, API Gateway, Lambda, CloudTrail, GuardDuty, DNSSEC.
Multi-account organisations rather than one account with everything in it. Smaller AWS builds too:
a serverless profile-photo checker
on Lambda + Rekognition, 14 Terraform resources.
AI agent security. The reason the projects above are built the way they are. I red-team tool-using AI agents in adversarial competitions: prompt injection, tool-call boundary abuse, and where an agent's guardrails actually hold versus where its own helpful behaviour leaks. Building agent infrastructure and attacking it turns out to be the same skill from two directions.
Security engineering as a habit. Five threat models across projects that did not require one. I have fixed ReDoS vulnerabilities CodeQL found in my own code, hardened FTS5 query construction against injection, and proved a "no network calls" guarantee by AST-walking every module rather than asserting it in the README.
Testing and verification. 2,000+ test functions written. One project ships a 41-gate verification harness that CI and local runs invoke identically, so the two cannot disagree about what "passing" means.
Infrastructure as Code. Terraform, GitHub Actions, hash-pinned dependencies, and PyPI publishing through OIDC trusted publishing rather than long-lived tokens.
Python Terraform Docker AWS Linux SQLite GitHub Actions pytest mypy Ruff
Bandit CodeQL MCP JSON-RPC MATLAB JavaScript
📍 Arlington, VA · 💼 LinkedIn · ✉️ yyaro@gmu.edu