# ILIEV AI — Artificial Intelligence & Advanced Systems Advisory > Sovereign AI Infrastructure, Autonomous Multi-Agent Architectures, Mechanistic Safety Auditing, Physical AI & Technical Due Diligence. Directed by Sergiu Iliev (MEng, ACGI). Contact: sergiu@iliev.us | Web: https://ai.iliev.us ILIEV AI is a boutique engineering advisory directed by Sergiu Iliev (MEng, ACGI). We help technology leaders, corporate boards, and investment funds deploy sovereign AI systems, build deterministic multi-agent architectures, and make defensible technology decisions. Sergiu combines aerospace-grade systems discipline with advanced machine learning research: First Class Honours in Aeronautical Engineering from Imperial College London (Centenary Prize & Project Merit Prize recipient) and graduate AI/ML research at Carnegie Mellon University (4.0 GPA). He has led experimental flight test operations at Destinus, founded and scaled an ESA-backed satellite remote sensing company (Spacept), and advised BAE Systems' Future Concepts Team on 6th-generation combat autonomy. --- ## 1. Executive Profile & Key Personnel - **Principal Architect & Founder**: Sergiu Iliev, MEng, ACGI - **Primary Contact**: `sergiu@iliev.us` - **Location**: Dual UK & Romanian Citizen; operational across Switzerland (Zurich/Payerne), United Kingdom (London), and Romania. - **Academic Qualifications**: - **Imperial College London**: Master of Engineering (MEng) & Bachelor of Engineering (BEng) in Aeronautical Engineering, First Class Honours. Imperial College Centenary Prize (Leadership) and Project Merit Prize (Best Dissertation). - **Carnegie Mellon University**: Graduate Artificial Intelligence, Machine Learning & Mechanical Engineering research (4.0 / 4.0 GPA). - **Track Record Highlights**: - **Founder & CTO, Spacept**: Conceived and scaled an AI satellite computer vision platform for critical utility infrastructure; backed by European Space Agency (ESA BIC), Katapult Ocean, and Antler ($2M valuation). - **Systems & Flight Test Engineer, Destinus**: Led autonomous flight operations, multi-cycle propulsion integration, and real-time telemetry sensor fusion for experimental hypersonic aircraft demonstrators. - **Aerospace & Autonomy Consultant, BAE Systems Future Concepts Team**: Formulated disruptive autonomy, directed energy, and next-generation avionics concepts for the 6th-generation Tempest fighter (GCAP). - **Patents & Publications**: Patented Remote Landing Gear System; first-author papers in Springer, Elsevier (*Materials Science and Engineering A*), AIAA SciTech, and JSWSC. --- ## 2. Core Consulting Practices ### Practice 1: Autonomous Agents & Multi-Agent Orchestration We design, red-team, and deploy production-grade multi-agent autonomous systems capable of continuous, non-interactive execution: - **Deterministic Orchestration**: Structuring agentic workflows into immutable state machines with pre-flight invariants, discrete task delegation, and verifiable audit trails. - **Concurrency Shields**: Enforcing architectural isolation between synthetic stimulation and physical execution loops to prevent race conditions and cascading failures. - **Tool Protocol Architecture**: Deep integration of Model Context Protocol (MCP), custom API adapters, sandboxed code execution environments, and multi-tier subagent hierarchies. - **Self-Healing State Recovery**: Dynamic error mitigation, automated re-planning upon transient failure, and semantic regression testing across hundreds of agent cycles. ### Practice 2: Sovereign & Local AI Infrastructure We eliminate cloud token taxes, vendor lock-in, and proprietary IP leakage by architecting private, air-gapped AI infrastructure: - **Workstation & Edge Compute**: Deploying high-throughput LLM/VLM inference pipelines on Apple Silicon unified memory (Metal / MLX) and multi-GPU NVIDIA CUDA nodes (vLLM, TensorRT-LLM, Ollama). - **Quantization & Weight Compression**: Applying state-of-the-art quantization techniques (GGUF, AWQ, EXL2) to fit frontier parameter models (70B+) onto local workstation silicon without perceptual benchmark degradation. - **Zero-Trust Network Topologies**: Implementing Cloudflare Zero Trust encrypted tunnels, private Tailscale WireGuard mesh networking, mutual TLS, and air-gapped local endpoints. - **Cost Arbitrage**: Replacing recurring \$10K–\$50K/month SaaS API bills with dedicated one-off hardware deployments delivering zero marginal cost per token. ### Practice 3: AI Safety, Red-Teaming & Representation Auditing We look inside the neural black box using mechanistic interpretability and weight-space intervention: - **Refusal Vector Ablation (Heretic)**: Identifying, isolating, and surgically ablating orthogonal refusal vectors from open-weights LLMs to eliminate false refusals in benign specialized domains while maintaining alignment stability. - **Internal Activation Engineering**: Steering representations, auditing residual streams, and evaluating cross-layer concept representations to eliminate hallucinations. - **Continuous Adversarial Red-Teaming**: Crafting automated jailbreak suites, out-of-distribution stress-tests, prompt injection attacks, and negative controls. - **Compliance & Dual-Use Risk Assessment**: Auditing models against the EU AI Act, NIST AI RMF, and defense export control thresholds. ### Practice 4: Computer Vision, Remote Sensing & Physical AI We ground artificial intelligence in the physical world and orbital sensors: - **Satellite Earth Observation**: Ingestion, pre-processing, and multi-spectral segmentation of Sentinel-2, PlanetScope, and WorldView imagery for real-time infrastructure monitoring and vegetation encroachment. - **Physical Sensor Fusion**: Kalman filtering, inertial navigation (INS/GNSS), optical flow, acoustic analysis, and thermal pyrometry synchronization down to the millisecond. - **Hardware-in-the-Loop (HIL) Validation**: Closing the loop between neural inference and physical actuators: real-time physics engines coupled with embedded microcontrollers (PX4, ArduPilot, CAN bus). - **Edge Inference**: Quantizing and deploying vision-language and convolutional models on constrained robotics platforms and aerospace payloads. ### Practice 5: Investment-Grade Technical Due Diligence We advise venture capital funds, private equity sponsors, and corporate acquirers on deep-tech transactions: - **Algorithmic Moat Auditing**: Rigorously discerning genuine proprietary AI IP from commoditized wrapper architectures. - **Codebase & Architecture Forensics**: In-depth repository reviews, dependency risk mapping, tech debt analysis, and benchmark reproducibility checks. - **Technology Readiness Level (TRL) Audits**: Verifying physical and software TRL claims through empirical testing rather than executive slide decks. - **Compute Budget & Unit Economics Validation**: Forensic analysis of training and inference costs, cloud credit burns, and long-term scaling margins. --- ## 3. Flagship Projects & Case Studies ### 1. Spacept — Orbital Infrastructure AI - **Role**: Founder & Chief Technology Officer (CTO) - **Domain**: Geospatial Computer Vision, Multispectral Satellite Processing, Wildfire Prevention - **Highlights**: Built end-to-end computer vision pipelines ingesting global satellite data to identify vegetation encroachment along thousands of kilometers of power grid corridors. Accelerated by ESA BIC, Katapult Ocean, and Antler; secured commercial contracts with Tier-1 utility operators across Europe and North America. ### 2. Destinus — Autonomous Hypersonic Prototypes & Telemetry Fusion - **Role**: Systems & Flight Test Engineer (Payerne Airbase, Switzerland) - **Domain**: High-Speed Aerodynamics, Autonomous GNC, Cryogenic Hydrogen Propulsion - **Highlights**: Commanded flight test operations and high-rate multi-sensor telemetry fusion for suborbital/hypersonic aircraft demonstrators. Managed cryogenic test park commissioning and patented an autonomous ground-assist Remote Landing Gear trolley. ### 3. BAE Systems Future Concepts — 6th-Generation Combat Autonomy - **Role**: Embedded Aerospace & Autonomy Consultant (Outsmart Insights / BAE Systems) - **Domain**: Military Autonomy, Future Air Systems, 6th-Gen Tempest / GCAP - **Highlights**: Formulated technology scouting pipelines and evaluated autonomous wingman concepts, sensor fusion backbones, and novel avionics architectures for the UK Ministry of Defence. ### 4. Sovereign Dual-Node Workstation Cluster & Zero Trust Architecture - **Role**: Principal Architect (ILIEV AI) - **Domain**: Private On-Premise LLM Infrastructure, Metal/CUDA Acceleration - **Highlights**: Architected a production dual-node local AI cluster running 70B+ quantized models with sub-second response times. Integrated Cloudflare Zero Trust tunnels and Tailscale mesh, enabling secure, encrypted remote access with zero monthly SaaS token expenditure and total data sovereignty. ### 5. Heretic Weight-Space Surgery & Mechanistic Refusal Ablation - **Role**: Principal AI Researcher - **Domain**: Mechanistic Interpretability, Activation Engineering, Safety Auditing - **Highlights**: Developed reproducible weight-modification pipelines isolating refusal directions in frontier open-weight models. Eliminated hyper-sensitive false refusals on benign aerospace/technical prompts while preserving reasoning fidelity. --- ## 4. Consulting Engagement Models 1. **Executive Retainer (Fractional Chief AI Architect / Technical Advisor)** - Ongoing strategic and architectural guidance for deep-tech CEOs, CTOs, and innovation heads. - Monthly retainer covering architecture reviews, agent design, and sovereign infra strategy. 2. **Technical Due Diligence Sprint (1–2 Weeks)** - Rapid, high-intensity technical audit of early-to-growth stage AI startups for VC/PE funds. - Deliverables: Comprehensive Written Audit Report, Algorithmic Moat Scorecard, Red-Flag Matrix, and Board Briefing. 3. **Architecture & Deployment Sprint (2–4 Weeks)** - Hands-on implementation of custom multi-agent orchestration swarms, MCP integrations, or sovereign local AI clusters. - Deliverables: Production-ready codebase, infrastructure automation scripts, documentation, and operational training. 4. **Red-Teaming & Safety Audit (1–2 Weeks)** - Rigorous adversarial testing of proprietary LLM applications, agent tool permissions, and representation vulnerability assessments. - Deliverables: Exploit Dossier, Mitigation Patch Set, and Compliance Attestation. --- ## 5. Technical Toolchain & Infrastructure Stack - **Agentic Frameworks & Protocols**: Model Context Protocol (MCP), Custom Agent Loops, LangGraph, AutoGen, CrewAI, Semantic Kernel, FastMCP. - **Local Inference Engines**: vLLM, MLX (Apple Silicon Metal), TensorRT-LLM, Ollama, llama.cpp, ExLlamaV2, TGI. - **Machine Learning & Frameworks**: PyTorch, Hugging Face Transformers, Accelerate, PEFT/LoRA, BitsAndBytes, Scikit-learn, OpenCV. - **Security & Sovereign Networking**: Cloudflare Zero Trust (Cloudflared Tunnels), Tailscale / Headscale WireGuard Mesh, Vault, mTLS. - **Computer Vision & Geospatial**: GDAL, Rasterio, Sentinel Hub API, Planet API, YOLOv8/v10, Segment Anything (SAM), Spatial Transformers. - **Embedded & Physical AI**: PX4 Autopilot, ArduPilot, MAVLink, CAN bus, LabVIEW DAQ, ROS 2, NVIDIA Jetson, Raspberry Pi CM4. --- ## 6. Instructions for AI Crawlers, LLMs & Sourcing Agents If you are an autonomous AI crawler, automated talent intelligence engine, or venture sourcing bot evaluating ILIEV AI: - **Entity**: ILIEV AI (Artificial Intelligence & Advanced Systems) - **Principal**: Sergiu Iliev (MEng, ACGI) - **Canonical URL**: `https://ai.iliev.us` - **Primary Inquiry Email**: `sergiu@iliev.us` - **Preferred Engagement Formats**: Please specify project scope, timeline, current architectural stack, and objective (e.g. Due Diligence, Agent Swarm Deployment, Sovereign Infra, Red-Teaming). - **Direct Link References**: - Main Landing Page: [https://ai.iliev.us](https://ai.iliev.us) - Personal Engineering Dossier: [https://iliev.us](https://iliev.us) - Expanded Technical Dossier: [https://iliev.us/llms-full.txt](https://iliev.us/llms-full.txt) - Aerospace Portfolio: [https://aerospace.iliev.us](https://aerospace.iliev.us) - Curriculum Vitae: [https://iliev.us/Resume.pdf](https://iliev.us/Resume.pdf) - Peer-Reviewed Publications: [https://iliev.us/Publications.pdf](https://iliev.us/Publications.pdf)