Most adaptive learning products reduce a student's understanding to a single percentage-correct number, which cannot separate a student who mastered a skill last month and is now rusty from one who never learned it. Modeling memory as something that truly decays over time, giving an AI the ability to see a student's handwritten work rather than only reading typed answers, and keeping a live multimodal AI conversation from silently drifting off course in a real tutoring session each required a different class of engineering.

The system also had to behave sensibly for a brand-new student with zero history, from the very first question, without waiting for training data to accumulate.

Each skill carries a continuous memory-strength value; the probability of a correct answer is a sigmoid over that strength minus the item's difficulty, and strength decays exponentially at a per-skill forgetting rate. Above the base model sit four deliberate design refinements: diminishing returns on repeated practice so grinding one skill cannot fake mastery, a response-time penalty so a slow correct answer counts as weaker evidence than a fast one, prerequisite back-propagation so a wrong answer also lowers confidence in every underlying skill it depends on, and a mastery freeze that stops decay once a skill crosses a real-mastery threshold.

For the second AI, the team built a supervisory service that ingests the live session feed over WebSocket and pushes instructions back into the tutor's live context over Server-Sent Events, for example detecting 60 seconds of silence and prompting the tutor to check in naturally, in its own voice. That service was deliberately re-architected from an early in-memory, single-instance design into a stateless, MongoDB-backed session model with automatic expiry, so the platform can scale horizontally across multiple concurrent classrooms.

Adaptive question selection, question loading, session and instruction orchestration, authentication, homework processing, payments, cost tracking, question generation, and an experimental avatar agent each ship as separate services, with a dedicated Pinecone vector store for long-term, cross-session student memory. The browser talks directly to Gemini's Live API over WebSocket rather than proxying audio through the backend, removing a full network hop of latency from every spoken exchange; because the raw API key cannot safely reach the browser, the team built a JWT-gated endpoint that mints a short-lived, use-limited ephemeral token instead, so the provider key never leaves the server.

Because live multimodal AI inference carries real per-token cost, the team also built cost governance into the platform from the start: usage is tracked per session, per modality, and per token class (including the 90 percent discount Google applies to cached content), with careful handling of the fact that the Gemini Live API reports token counts cumulatively rather than incrementally, a subtlety that would otherwise silently inflate reported cost many times over if missed. A four-tier model router sends simple requests to cheaper, faster models and only escalates to premium-tier inference when task complexity actually warrants it.