← TKF talk · Slide 2 of 18
The problem
Needlestick injury. A clinician needs the exposure protocol right now - not 15 minutes from now, not after hunting through PDFs or finding a binder.
This data is PHI-adjacent. Even though our policies aren't patient records, they sit right next to them. If cloud AI becomes the default today, it stays the default when real PHI shows up tomorrow. We have to do this right - and that means local AI.
A single PHIPA breach: $1,000,000 in court, or $500,000 Privacy Commissioner fine. The liability is on us, not the vendor.
CLOUD AI WON'T WORK HERE
- 📋 PHIPA: if the vendor leaks our data, liability falls on us.
- 🧠 Generic AI hallucinates: not trained on our docs, makes things up.
- 📤 Custom GPTs / Claude Projects: still upload our docs to the cloud.
- 🔧 Fine-tuning failed: six runs, still hallucinated. Every doc change
- meant a full retrain.
SOLUTION: RAG PROXY
- 🔍 Reads our actual documents: searches and pulls the right passage
- before the AI answers.
- 💬 Quotes, doesn't guess: answers come straight from our documentation.
- 🏛️ Lives on our server: nothing ever touches the cloud.
- 📝 Edit a doc, done: next answer reflects the change. No retraining.