News

27 Aug 2026. Faculty AI Kitchen lightning talk. It went well. Same RAG Proxy story: clinic answers on Faculty hardware, not in a cloud chat app. Two NVIDIA DGX Sparks are in the building. Pins (cheat sheets the proxy injects) and a software harness (policy-number check) are in production. The Perl files on this site are still the first public snapshot. See the code note.

RAG Proxy

RAG Proxy

Local AI that knows our policies, and keeps our data inside the building

Faculty of Dentistry · University of Toronto

Talks

🎬

TechKnowFile

May 2026. The long deck. Eighteen slides, picture first, talking notes underneath.
🍳

AI Kitchen

27 Aug 2026. Faculty jam. Five slides, notes under each, plus the recording. It went well.

All talks →

87%
RAG Accuracy
50.4%
Fine-Tuning
$0
Monthly Cost
100%
Local & Private

How it works

⚙

RAG Setup

The full technical architecture: vector search, proxy interception, and the code.
📌

Pins

Cheat sheets the proxy injects. Not vector search.
🛡️

Harness

Software that rejects invented policy numbers.
📚

Glossary

AI and RAG terms explained in plain English. No jargon, no hand-waving.

About

📖

The Story

How six failed fine-tuning runs led to a 730-line Perl solution that actually works.
❌

Six Failed Runs

What was tried, what went wrong, and what each failure taught us. Not a talk. The write-up.
📅

Changelog

The full timeline from first fine-tuning attempt through AI Kitchen.
🔗

Resources

Research papers, tools, and references from the talks.
See the talks →

Michael Limberger · Senior Web & Linux Administrator · Faculty of Dentistry