Resources

The proxy injects a known-good file for a topic. Not a search score. Used when retrieval might miss.

Software after the model answers. Invented policy numbers fail. The model has to redo the reply.

27 Aug 2026 Faculty talk. Recording plus five slides with the delivered script underneath.

First generation model. Mistral's 14B instruct checkpoint with a vision encoder. Ran locally as ministral-3:14b (Ollama) for the TechKnowFile snapshot. Client name dental/ministral-3.

Next step, on Faculty Sparks. Hybrid MoE, about 30B total / 3B active. Client name dental/nemotron-3-nano.

Current generation model. 30B-A3B NVFP4 on Sparks (vLLM). Successor to Nano. Client name dental/nemotron-3.5-lightning.

The Faculty Sparks. Dell's branded NVIDIA DGX Spark: Grace Blackwell GB10 Superchip, 128 GB unified memory, DGX OS. Two of these boxes run generation and the clinic chat path on premises.

NVIDIA's own GB10 desktop. Same Grace Blackwell chip as the Dell Pro Max. What people mean when they say Spark.

Microsoft, EMNLP 2024. The head-to-head comparison that showed RAG at 87.5% accuracy vs fine-tuning at 50.4%. The paper that validated our pivot.

Allen-Zhu & Li, ICLR 2025. Proved that each fact needs 100-1,000 training exposures to be reliably learned. We gave ours about 15. Mystery solved.

Zhou et al., 2023. Showed that fine-tuning primarily teaches style and format, not new knowledge. The model learns how you talk, not what you know.

Stanford, 2024. Found a 37% ceiling on new knowledge absorption through fine-tuning even with commercial APIs. The problem isn't technique, it's approach.

SMU/Microsoft, 2024. A hybrid fine-tuning + RAG approach. Interesting research, but pure RAG proved simpler and sufficient for our use case.

Run large language models locally. The engine behind our AI inference. Download a model, run it, talk to it through an API. That simple.

A self-hosted chat interface for AI. This is what our clinicians actually interact with. Clean UI, conversation history, file uploads.

A real-time web framework for Perl. The backbone of RAG Proxy. Handles HTTP interception, async requests, and WebSocket connections.

Fast numerical computing for Perl. Does the vector math: cosine similarity calculations across thousands of document embeddings in milliseconds.

A serverless database engine. Stores our document chunks and metadata. No daemon, no config, no port. Just a file. Perfect for our scale.

BAAI's multilingual embedding model. Converts text into 1,024-dimension vectors. Runs locally through Ollama. Handles English, French, and mixed-language queries.

A tiny SQLite wrapper for Mojolicious. Gives us migrations, connection pooling, and a clean query interface. The glue between Brain.pm and the database.

University of Toronto's annual IT knowledge-sharing conference. Where this talk is being presented, May 6-7, 2026.

Canada's largest dental faculty. Home to the RAG Proxy system, where clinicians and staff use it daily for policy lookups.