Tule
- AWS Bedrock
- RAG
- Titan Embeddings
- Pinecone
- React
- Node.js
Tule is a campus information platform for UC Merced: students ask a question and get a structured answer card, not a wall of chat text. I scoped it, shipped it on AWS with a RAG pipeline, and wired it into DineBoard so live dining menus are public the moment staff update them.
The problem
Campus information lives in dozens of places — dining hours, menus, services, logistics — and students mostly find it by asking someone or giving up. The bet behind Tule: one place to ask, with answers structured enough to trust.
Product decisions
The defining choice was structured response cards over a chat transcript. Students submit a query and receive a card built for the answer — hours, locations, menus — which keeps responses scannable and keeps the RAG pipeline honest about what it knows.
The second was integration over content: rather than copying menu data into Tule, I connected it directly to DineBoard. Any station's live menu is accessible in real time, up-to-the-minute with whatever staff pushed through the dashboard.
The build
Tule is AWS-hosted end to end. Queries run through a retrieval-augmented generation pipeline built on Amazon Bedrock with Titan Embeddings, backed by Pinecone for vector search. The pipeline retrieves campus content, grounds the model's answer, and renders it into a typed response card.