A working argument for what an adaptive AI tutor should be. A learner works a lesson with an AI tutor while a live inspector shows the whole pipeline — behavior → xAPI → inferred state → adaptation — with every inference shown honestly, including its confidence (capped below 100% on purpose) and what drove it. The learner can see, contest, and override what the system infers.
The core design bet: an adaptive system should never assert more confidence than its evidence supports, and the learner should always be able to see and challenge its reasoning. The live inspector shows the exact behavioral signals — pacing, hesitation, review patterns — that produced each inferred state, before any adaptation happens. Built solo with Claude as production partner across the full stack: React front end, an xAPI event pipeline into a Learning Record Store, and an inference layer translating raw behavior into pedagogical decisions.
Building the prototype opened onto a bigger thesis: most "adaptive" tutors reset with every session. The interesting problem, especially for corporate L&D, is a persistent adaptive tutor — a learner-tutor relationship that builds a profile over time instead of starting cold each time someone opens it. That thesis split into three research threads I've kept developing past the prototype itself.
Thread A — Persistent Adaptive Tutor. Cognitive load is the easier half to infer: pacing, error rates, time-to-completion, how often someone asks for clarification. The harder half is behavioral and contextual — when a person learns best, whether they need a concrete example before an abstraction or the reverse, whether they thrive on productive struggle or get frustrated by it. That's the same territory Josh Bersin's Dynamic Enablement model points at, and it's where a persistent profile actually earns its keep.
Thread B — Privacy-Preserving Adaptive Tutoring. Enterprise data-minimization policy is a real design constraint, not a footnote: a tutor built for a corporate context has to infer learning style and cognitive load from implicit behavioral signals, with minimal explicit data collection. Most of the literature treats privacy and personalization as a trade-off to be balanced — I'd argue that's a genuine research gap, not a settled question. The technical path runs through federated learning (FedProx has shown gains over FedAvg specifically on heterogeneous student data), differential privacy, on-device data locality, and explainability techniques that keep the inference auditable instead of opaque.
Thread C — Intent-Aware Learning Design. Not every question deserves the same kind of answer. Sometimes a learner needs a fast, direct answer to finish a task under deadline pressure; sometimes what they actually need is for the knowledge to stick. The system distinguishes a direct-answer mode from a retrieval-practice mode, and flags when a pattern of repeated quick-answer requests might itself be the signal that deeper learning — not just task completion — is what's needed.
Supporting research runs alongside all three threads: NSF-funded work out of AI-ALOE at Georgia Tech, RCT evidence that AI tutoring can outperform traditional instruction in some studies, and federated learning as the concrete technical path for the privacy-preserving angle.
Try the prototype →
