-4ade80?style=flat-square&labelColor=020b04&color=14532d){kind=icon}
Independent researcher and developer building production AI agent infrastructure — security frameworks, persistent memory systems, autonomous reasoning agents, and the tooling that connects them. Self-taught developer with 8+ years of prior experience in emergency medicine, where I learned what "production reliability" means in environments where failure has immediate consequences.
I didn't come to AI through a CS degree. I came because I saw problems that needed solving, taught myself to build the tools, and haven't stopped. The systems I've built are in production, tested, and doing real work.
AI Agent Security Framework — obol
Capability-based access control for autonomous AI agents. Core concept: tools don't exist until a scoped plan is approved by a watcher AI. Agents operate under principle of least privilege with runtime enforcement.
- 96,000+ lines of Python with 107+ tests across 27 test files
- 8-layer defense: MCP proxy, command control, capability profiles, AI-evaluated plans, inbound content scanning, tripwire honeypots, API key proxy, kernel audit (auditd/iptables)
- Pre/post-tool hooks compatible with Claude Code's hook system (PreToolUse/PostToolUse)
- AI watcher evaluates proposed plans against capability profiles (~$0.01/evaluation via Claude Haiku)
- Live in production — actively securing agent workloads
Cross-Model Episodic Memory — engram
Persistent memory system that works across any LLM. Hook-based architecture fires on every prompt in under 1 second.
- Entity extraction via lightweight LLM (Haiku) or zero-cost regex-only mode
- Ranked recall: recency x frequency x explicit feedback x co-occurrence
- Fuzzy entity matching (Levenshtein distance), capped retrieval to prevent context flooding
- Zero external dependencies, SQLite-backed, ~$3/year operational cost
- Model-agnostic — works with Claude, GPT, Gemini, local models
Local Autonomous Reasoning Agent — gor-agent
ReAct-loop agent running entirely on local LLMs via Ollama. Multi-step autonomous reasoning with SQLite state persistence, session management, and cross-session history.
- 9 built-in tools: file operations, shell, network diagnosis/repair, malware analysis, system diagnostics, scratchpad
- 30-step execution limit with working memory
- 16-scenario automated test harness with scoring database
- Tested with Qwen 2.5 14B and Gemma 4 27B on consumer GPU (AMD RX 9070 XT / ROCm)
- Zero cloud dependency — runs fully local
LLM Compression Language — strix
Token-aware compression for LLM context windows. Exploring whether a purpose-built compression language can meaningfully extend effective context. Learned pattern mining from real conversation data. Targeting 2-3x compression at >95% semantic fidelity. Research stage.
Built 15 Model Context Protocol servers covering: persistent memory, penetration testing orchestration (50 tools), home automation + network security (10 tools), router/WiFi administration (18 tools), browser automation, budget integration, device management, AI search, telemetry, and more. All MCP stdio architecture with JSON-RPC.
API gateway with auth routing, AI-powered news curation via Workers AI, SMS gateway (Twilio integration), webhook dispatcher with schema validation, autonomous data collection agent. ~4,800 LOC TypeScript on Hono framework.
Multi-provider cloud provisioning system spanning Hetzner, Kamatera, Vultr, and Vast.ai. Automated lab spin-up/teardown for security research workloads. Cloud-init based, SSH key managed, full lifecycle automation.
PyQt6 desktop application for operational management — tabbed interface (F1-F5), command palette (Ctrl+K), integrated agent orchestration, recovery system. Wayland-native on KDE Plasma.
| Area | Details |
|---|---|
| AI Alignment | Adversarial fine-tuning, backdoor implantation analysis, RLHF poisoning, model merging attacks, representation engineering. Responsible disclosure only. |
| Prompt Injection | Systematic injection testing across model families. Multi-step and goal-aligned attack vectors. Baseline measurement methodology. |
| Vulnerability Research | AFL++ fuzzing with ASAN, automated crash triage (CASR clustering), crash-to-PoC pipeline. Elastic compute across 4 cloud providers. |
| Penetration Testing | Built structured 7-phase pentest orchestration with 50+ tools, evidence capture, browser automation, and AI-powered peer review. |
| Category | Details |
|---|---|
| Languages | Python (primary, 100K+ LOC across projects), TypeScript, Bash/Fish, SQL |
| AI/ML | Claude API (Anthropic SDK), Ollama, PyTorch (ROCm), Whisper, YOLO, Flower (federated learning), prompt engineering |
| Agent Tooling | MCP protocol (server + client), ReAct loops, tool-use architectures, hook systems, capability-based security, plan-evaluate-execute |
| Infrastructure | Cloudflare Workers, Docker, systemd, cloud provisioning (Hetzner / Vultr / Kamatera / Vast.ai) |
| Databases | SQLite (heavy — used as inference engines and state machines), PostgreSQL, session state management |
| Security | AFL++, ASAN, auditd, iptables, adversarial ML, penetration testing, vulnerability research |
| Systems | Linux (daily driver), KDE Plasma, Git/GitHub, FastAPI, Hono, ROCm GPU compute (AMD) |
Worked across urban trauma centers, rural critical access hospitals, tribal healthcare (IHS), crisis stabilization, and float assignments spanning orthopedics, PACU, cath lab, and behavioral health. All major EMR systems (Epic, Cerner, Meditech, Medhost).
Why this matters for AI:
- Built intuition for high-stakes, real-time decision systems where failure is immediate
- Deep understanding of AI deployment in regulated, safety-critical environments
- Experienced with human-system interaction friction across multiple documentation platforms
- Domain expertise valuable for healthcare AI, clinical documentation, and safety-critical agent systems
Associate Degree in Nursing — 2016 Certifications: BLS, ACLS, PALS, TNCC, ENPC (all current) Self-taught developer — Python, systems programming, AI/ML, security research (2+ years intensive, full-time equivalent)
Roles in AI agent development, AI infrastructure/operations, or AI security research. I'm drawn to teams building autonomous systems, agent tooling, or AI safety infrastructure.
I bring production agent systems I built and operate, security research depth, and the perspective of someone who's worked in environments where system reliability is life-or-death — not a metaphor.
Auto-generated PDF available in Releases