Built a fully local, multi-agent AI coding engine. OpenCode serves as the dev interface. LangGraph runs the control orchestrator. Graphify maps the codebase knowledge graph. The entire stack runs on dual RTX 3090 GPUs via Ollama, cycling Qwen3-Coder 30B for code generation, Qwen3.6 27B for parallel review, and DeepSeek-R1 32B for adversarial red-teaming. Hardened by OpenCode skills — Superpowers, Caveman, Grill Me, Handoff, Firecrawl, and Understand-Anything — to enforce bulletproof coding methodology with zero cloud API dependencies.

Built and deployed OpenSOC, local-first SOC agent that ingests firewall syslog into a SQLite property graph engine, reasons over threats with a local LLM via llama.cpp, and contains them through the firewall's REST API. Added SOCbot, a two-way interactive communicator over Discord (with email/SMS as pluggable future channels) supporting allowlisted operator commands (!soc ban, status, blocks, unblock) plus a daily automated briefing. A modular connector framework ships with built-ins for firewall, SIEM, threat intelligence, and GeoIP enrichment service, with custom user-written connectors supported. The pipeline is fully automated with a human in the loop: nothing is blocked without an allowlisted approval through SOCbot and guardrails.

Self-hosted GPU infrastructure for production AI inference and contract analysis. Dell R730 with dual RTX 3090, Proxmox hypervisor, vLLM inference engine, FastAPI orchestration, PostgreSQL/Redis/Qdrant backends.

Proof-of-concept AI system to extract key clauses and generate risk summaries from legal contracts. Implemented REST APIs to handle PDF uploads, parse content, and return structured JSON outputs.

Built a semantic search platform for enterprise documents with per-tenant isolation, embedding-based search, and access control via RBAC.

Automated audit logging and tenant-specific access reporting using AWS CloudTrail, CloudWatch, and AWS Config. Deployed repeatable security controls with Terraform and CloudFormation.