Agentic AI Systems
ARKONA — Autonomous AI Ecosystem
A 6-domain production ecosystem with 47 services and 18 AI agents operating autonomously around the clock. Domains span hardware reverse engineering, firmware analysis, software development, intelligence collection, enterprise operations, and data management. Agents follow a daily battle rhythm — research at 0200, build at 0300, publish at 0615, brief at 0645.
Key Challenges Solved
- Agent harness design with persistent memory, context compression, and structured communication architectures across 18 agents
- Longer-horizon autonomous operations (overnight research, build, publish cycles) with thermal monitoring, circuit breakers, and self-healing
- Rigorous prompt engineering and evaluation pipelines across heterogeneous LLM backends
- 47K+ lines of production code across Python, JavaScript, Shell, and YAML
Multi-Agent
Production
Autonomous
6 Domains
COMET — AI Governance Framework
Cognitive Operations and Mission Effectiveness Taxonomy. Takes any domain — aircraft maintenance, business operations, cybersecurity — and maps every role and task, then classifies each across five delegation levels from fully human to fully autonomous. All grounded in 20 industry standards (NIST, ISO, OWASP). Outputs a RACI matrix where AI agents sit alongside human roles as first-class participants.
Key Challenges Solved
- Quantitative benchmarking framework for agentic task delegation across any operational domain
- Automated evaluation of agent performance against 20 industry standards with auditable scoring
- Facilitated workshop flow for real-time stakeholder consensus on delegation levels
Governance
AI Safety
Standards
RACI
MuXD — Hybrid LLM Router
Intelligent routing layer that classifies every AI request and decides whether it needs cloud intelligence or whether a local model can handle it. Makes thousands of routing decisions per day across multiple model backends. Achieves 47% API cost savings while maintaining output quality on complex tasks.
Key Challenges Solved
- Task classification and prompt optimization for accurate routing decisions across heterogeneous models
- Graceful fallback chains with automated evaluation of output quality per model
- Dual-GPU load balancing with thermal-aware scheduling
LLM Routing
Cost Optimization
47% Savings
Production
FORGE — AI Software Factory
Multi-agent development environment where 7 specialized AI agents collaboratively plan, implement, test, and deploy software. Includes a Skill Builder pipeline that imports AI-native tasks from COMET governance, auto-collects training data from agent execution, and fine-tunes local models via QLoRA — graduating tasks from cloud inference to cost-optimized local routing through MuXD.
Key Challenges Solved
- Closed-loop governance-to-inference pipeline (COMET → Agent SDK → Fine-Tune → MuXD)
- Inter-agent coordination for code review and quality gates
- QLoRA fine-tuning on dual P40 GPUs with automated training data curation
Software Factory
7 Agents
Skill Builder
Fine-Tuning
Local Model Fine-Tuning Research
Applied research into task-specific fine-tuning of open-weight foundation models on consumer-grade GPUs. Built an automated pipeline from training data curation through QLoRA fine-tuning to model deployment, enabling governed transition of agent tasks from cloud inference to on-premise execution at near-zero marginal cost.
Key Challenges Solved
- QLoRA 4-bit fine-tuning on Pascal-generation GPUs (2× P40) — no Ampere/Volta required
- Multi-GPU training via DeepSpeed ZeRO-3 for models exceeding single-GPU VRAM
- Automated training data pipeline: agent execution logs → schema validation → quality labeling → deduplication → JSONL export
- Model evaluation and governance-controlled graduation from cloud to local inference
Fine-Tuning
QLoRA
DeepSpeed
Llama / Phi / Gemma
On-Premise AI
SCHOLAR — PhD Study Platform
AI-augmented research and study platform with 12 functional tabs, SM-2 spaced-repetition flashcards, and a 5-agent research pipeline. Includes AI chat for on-demand concept exploration and automated literature review agents that scan publications daily.
Key Challenges Solved
- Integrating spaced-repetition algorithms with AI-generated content
- Multi-agent research pipeline with source quality evaluation and relevance scoring
- Maintaining coherence across daily research runs with persistent memory
Research
Long-Horizon
Memory
Education
VAULT — Evidence Management System
Digital evidence vault for cyber-physical reverse engineering artifacts. Integrates with Wiki.js for automated publishing, includes a KiCad agent that converts PCB photographs into engineering schematics using AI vision, and maintains chain-of-custody metadata for all stored evidence.
Key Challenges Solved
- AI vision pipeline for PCB photo to KiCad schematic conversion
- Automated evidence cataloging with chain-of-custody tracking
- Wiki.js integration for publishing research findings
Evidence Mgmt
AI Vision
KiCad
Reverse Engineering
Cybersecurity & Defense
Cyber Physical System (CPS) Hardening
Leading cross-functional cybersecurity teams in the pursuit of cyber hardening Cyber Physical Systems against nation-state threats. Directing both offensive and defensive cyber system engineering capabilities against USG priority systems at Percival Engineering.
Key Challenges Solved
- System threat analysis and vulnerability assessments on operational technology
- Formulating strategies to fortify weapon systems against cyber intrusions
- Bridging OCO and DCO engineering capabilities for holistic CPS defense
OCO/DCO
CPS
Threat Analysis
Current
Weapons & Space Cybersecurity Labs
Provided technical leadership and direction to the NSA Weapons and Space Cybersecurity workforce and laboratories — 250+ employees and $36.8M budget. Served as Chief Cybersecurity Engineer and Deputy Military Technical Director for the 7th Intelligence Squadron.
Key Challenges Solved
- Scaling cybersecurity lab operations across classified environments
- Technical direction for weapon system vulnerability research
- Workforce development for specialized cyber engineering talent
NSA
Lab Operations
250+ Personnel
$36.8M
NSA CNODP — Computer Network Operations
Graduate of the NSA's 22-week PhD-level Computer Network Operations Development Program, followed by 30 months of technical tours. Training covered advanced software analysis, reverse engineering, network security, cryptography, programming, Windows/Linux system internals, and operating system security.
Key Challenges Solved
- Advanced binary analysis and reverse engineering of complex systems
- Network exploitation techniques and defensive countermeasures
- Operating system internals for vulnerability research
NSA
Reverse Engineering
Cryptography
CNODP
Cyber System Risk Analysis (CSRA) Methodology
Co-developed the CSRA methodology while directing a 63-member flight of engineers and analysts conducting cybersecurity tests of DoD aircraft and weapon systems. Subsequently led the stand-up of AFSOC/SOCOM's first Operational Test Cyber Flight at the 18th Flight Test Squadron.
Key Challenges Solved
- Standardizing cyber risk assessment across diverse weapon system platforms
- Standing up operational test capabilities for aircraft cybersecurity from zero
- 1st Place Spark Tank Innovation (USAF/AFSOC MAJCOM) for ‘benign malware’ concept
Methodology
Test & Evaluation
AFSOC
Innovation Award
Non-Kinetic Counter Electronics
Led multi-discipline government and industry team conducting weapon lethality assessments of Air Force non-kinetic weapons at AFLCMC. Evaluated electromagnetic effects on target electronics and developed assessment frameworks for directed energy systems.
Key Challenges Solved
- Weapon lethality modeling for non-kinetic effects on electronic systems
- Cross-discipline coordination between government and industry teams
- Assessment frameworks for emerging directed energy capabilities
Directed Energy
AFLCMC
Lethality Assessment
Electronic Warfare
These projects represent systems designed, built, and deployed across 25 years of cybersecurity and AI systems engineering. Happy to discuss architecture decisions, challenges, and lessons learned in detail.