ai systems / Engineering project

Agent Infrastructure & AI SOC

Multi-agent workflows, scoped tool permissions, and local GPU inference bench

Engineered multi-agent systems with per-role tool allowlists and intelligent model routing, powered by a self-hosted GPU inference bench (dual RTX 5070 Ti + RTX 4070 SUPER running llama.cpp and CUDA). Built a 19-adapter personal data pipeline (PostgreSQL, pgvector, local embeddings) feeding agent long-term memory. Operates a security laboratory where a strictly guarded Python agent is the sole entity allowed to modify firewall rules, with comprehensive injection testing and telemetry capture.

Architecture & decisions

• Self-hosted GPU Bench: Low-latency local inference with quantization and strict VRAM budget management.
• Scoped Tool Sandboxing: Agents operate under strict JSON schemas with deterministic output validation.
• 19-adapter ETL Memory Pipeline: Feeds contextual embeddings into pgvector vector store.
• Security Lab: Suricata and nftables integration tested against prompt injection and privilege escalation.

Technology

What to explore in a walkthrough

• Where the system runs today and which parts are production, development, or research.
• The architectural tradeoffs, failure modes, and tests behind these decisions.
• What AI agents contributed and how their output was reviewed.
• Operational evidence and measurable results available for review.

Project descriptions and experience are candidate-provided. This résumé's MCP tools are directly testable; other systems are available for an architecture walkthrough. Current and past employment positions and employers are private. Professional focus and target roles do not disclose a current job. Independent verification and references require direct contact.