How the learning companion thinks
Three animated views of the same system: the companion brain as a hub-and-spoke, the full Tony substrate that powers it, and the retrieval architecture that grounds every answer in SFU's own course content.
The SFU Companion Brain
Six signal streams feed the central companion continuously. Every student interaction on any stream makes the whole loop more useful. This is the self-reinforcing mechanism that separates a real companion from a bolted-on LMS feature.
Tutor Chat + Retrieval
Grounded in SFU's own module content. Every answer is drawn from real course material, not the open web.
Quiz and Summary Engine
On-demand quiz generation, audio summaries, and flashcard sets from any module or uploaded reading.
The Self-Improving Loop
Student signals feed back into the companion. Paths that help students learn get stronger. This is Tony's substrate edge.
Tony's Substrate: Three Agent Layers
Tony's one.ie platform is not a single AI chatbot. It is a three-layer agent architecture. The education-specific companion sits inside the Specialist layer; the whole stack runs on production-grade infrastructure that Oakley does not need to fund an in-house team to build.
What this means for SFU
The SFU Study Companion is a Specialist Agent sitting inside an enterprise-grade orchestration platform that already exists. OO builds the education-specific companion layer and the pedagogy. Tony's team maintains the infrastructure, the connector layer, and the substrate. SFU gets the outcome without building any of the plumbing.
Why every answer is grounded, not hallucinated
The study companion never answers from the open web. It retrieves from SFU's indexed course content using a RAG (Retrieval-Augmented Generation) pipeline. This is what makes an AI tutor safe and credible in a health university context: Medicine, Psychotherapy Science, Psychology, and Law all require accurate, source-grounded answers.
Why RAG matters for a health university
When the subject is psychotherapy, medicine, or law, a hallucinating AI is not a minor inconvenience - it is an academic liability. A RAG pipeline that retrieves from SFU's own indexed, approved course material means every answer is grounded in material the faculty already stands behind. The LLM synthesises; it does not invent. This is the credibility layer that makes an AI tutor acceptable to clinical academics.
IU's Syntea uses the same architectural principle. It is not a general-purpose chatbot. It is a grounded companion that retrieves from IU's course content. That is why the outcomes are measurable: 27% study-time reduction is achievable because the answers are consistently accurate.