High-Fidelity Infrastructure for High-Integrity Learning
It’s not a wrapper – it’s a RAG (Retrieval Augmented Generation)
CustomAILab does not utilize "wrapper" logic. We have engineered a federated, multi-agent RAG (Retrieval-Augmented Generation) environment designed to solve the four core educational failings of standard LLMs: Hallucination, Academic Integrity, Sycophancy, and Data Privacy.
1. Federated RAG & Semantic Chunking
Unlike generalist AI, Cal utilizes a Local Knowledge Base with vector embeddings stored in pgvector, orchestrated by a Canadian-hosted Google Cloud Run (GCR) instance.
The Tech: All documents in Cal’s Knowledge Base are run through Gemini Vision to produce clean, readable markdown. They are then split into large semantic ‘chunks’ based on unit/chapter headings before they are embedded as vectors in the knowledge base.
The Benefit: This means Cal retrieves the specific section of a textbook or course document that a question is about, rather than guessing from general training.
2. Multi-Agent Orchestration
Our system operates through three distinct LLM instances that serve as "checks and balances" for one another:
Cal (The Tutor): Optimized for socratic dialogue and pedagogical agility. Cal uses a system of "Context Injection" to adapt his personality to the student's chosen learning style (e.g., Step-by-Step vs. Analogy-Based).
The Librarian (The RAG System): Drawing on a shared universal knowledge base consisting of Ontario curriculum documents and OER textbooks, as well as the specific course documents uploaded by the student into their own personal knowledge base, the Librarian provides Cal with the information he needs to answer a student’s query.
The Dean (The Auditor): A separate, independent LLM auditor. The Dean does not interact with the student; it reviews chat logs against our proprietary Cognitive Depth Index (CDI) to measure student engagement and the depth of understanding. These reviews are summarized in a weekly report emailed to the parent.
3. Advanced OCR & STEM Notation
To support rigorous high school STEM curricula, we have integrated Gemini Vision Document Intelligence.
The Tech: This allows the conversion of complex PDFs, chemical diagrams, and handwritten equations into high-fidelity LaTeX and Markdown.
The Benefit: Cal can "see" and solve math and physics problems with high symbolic accuracy. This includes a student’s handwritten notes, assignments and tests.
4. Data Residency and Compute
All student data — accounts, chat logs, uploaded documents and reports — is stored in Montreal (Supabase and Google Cloud Run, northamerica-northeast1). The AI models themselves currently run on Google's and Anthropic's enterprise cloud endpoints; we will pin model inference to the Montreal region as soon as the providers make it available there, and we will say so on this page when it is done. Both providers are accessed through enterprise API agreements under which student lessons are never used to train their models.
Encryption: All student interactions are encrypted at rest and in transit.
Enterprise API: Our exclusive use of enterprise API keys ensures that lessons cannot be used for LLM pre-training or for any other commercial purpose.