ls projects/ --sort=impact
USAP — Unified Security Agent Platform
79 cybersecurity skills and 12 orchestrator agents exposed over the Model Context Protocol. Every verdict is grounded, typed, and auditable.
- —Enforced resolvable-evidence gate — no verdict without a source that resolves
- —Typed 11-field output contract across every skill
- —Human-approval gating on every mutating action
- —Tamper-resistant, hash-chained audit log
ARIA
7-agent architecture driven by a locally fine-tuned model (Llama 3.1 8B, 181k-sample training corpus, MLX), in evaluation phase against DVWA and OWASP Juice Shop in an isolated lab. Not deployed against production systems.
Private research, evaluation phase - write-up on request
Active investigations
ps aux | grep building
Secure LLM inference pipeline on AWS
Serving an open model on AWS hardened end-to-end: Terraform IaC, least-privilege IAM, private networking, secrets management, CI security gates, and prompt-injection guardrails.
Cloud-security guardrails as code
Terraform baseline with Checkov and OPA/Conftest policy-as-code, a CI gate that blocks violations, and drift detection.
Agentic-AI red-team lab
garak, PyRIT, and promptfoo attack suites run against an LLM agent application, with findings mapped to OWASP LLM Top 10 and MITRE ATLAS.