An Intro
Hey everyone — J.C. Marin, Melbourne, FL. Technical Director of AI Engineering at Provolega, where I lead a small team shipping agentic systems into production in a regulated environment.
Before that, aerospace/defense at Northrop Grumman, where I built a RAG system for the hardest environment there is: air-gapped, locked-down machines — no internet, no installs, nothing leaves the box. One command, fully local: hybrid semantic + keyword retrieval, offline OCR, cited answers, and an eval harness so retrieval changes get scored instead of guessed at. It's public now as 1clickrag.
After hours I build FrontierHarness, a project-first desktop workspace for running multi-agent coding workflows (Claude, Codex, OpenCode) with human-controlled merges, and LogansOS, a local-first AI-native OS. I also wrote Building AI That Knows Your Data: RAG From Scratch, and I run local models on my own GPUs — llama.cpp, quantized MoEs, no telemetry.
Here to compare notes with people actually shipping agents: orchestration patterns, evals, retrieval, and where multi-agent setups break in the real world.