Primary speaker:
Michael Limberger, Senior Web & Linux Administrator, IITS, Faculty of Dentistry
Description:
Staff across the Faculty of Dentistry need fast, accurate access to 94 operational policy documents covering everything from needlestick injury protocols to controlled substance procedures. The existing solution? A shared drive full of PDFs. During a percutaneous exposure, nobody wants to browse a file server looking for policy H.03.
This is the story of what I built, what failed and what I learned.
I started with fine-tuning: training a local AI model directly on the policy documents. I built ModelForge, an automated pipeline on Apple Silicon, and ran six training runs across three architectures. The models learned to sound professional, but fabricated email addresses, invented form names and eventually refused to answer the very questions they were trained on. The peer-reviewed research confirmed my findings: fine-tuning teaches style and format, not factual knowledge.
I pivoted to Retrieval-Augmented Generation (RAG) and built RagForge: a pipeline that takes a folder of documents and deploys a working AI knowledge assistant in one command. RAG worked. But the search layer underneath wasn't mine. When it returned the wrong policy, I couldn't fix it. So I replaced it entirely.
RAG Proxy sits between the user and the AI. Every question passes through it, my own retrieval pipeline runs first, and the AI answers from what it finds. The client sees nothing different. The answers just get better. No other tool does this transparently.
Every answer cites its source document by policy number. The entire stack is open-source and locally hosted. No institutional data ever leaves our infrastructure.
This is a proof of concept built by one administrator with no budget. The goal is to demonstrate what's possible so the Faculty can invest in a production-grade version serving staff and students across our clinics and programs.
This session is a lecture with live demo and Q&A. Attendees will see a real query answered with sourced citations, a hallucination resistance test, an out-of-scope refusal and the one-command deployment pipeline in action.
