Technology
Six layers, each with a narrow job. The interesting engineering is in the constraints — keeping a generative system grounded, measurable and interruptible.
A curriculum graph of concepts, prerequisites, objectives and assessment items, aligned to board and grade. Retrieval is scoped to this graph, which is what keeps generated explanations on syllabus.
Language models with tool access, running against the knowledge layer. Tools let the tutor drive simulations, pull assessment items, and check the learner record before deciding what to say next.
Bayesian knowledge tracing maintains a probability of mastery per concept, updated from answers, speech attempts, hesitation and applet interactions. Prerequisite gaps surface before they compound.
Streaming voice in both directions with sub-second turnaround, so a lesson feels like a conversation rather than a request-response cycle. Session state survives reconnects and device switches.
Pronunciation assessment scoring accuracy, fluency and completeness down to the phoneme, plus generated reference audio so a learner can hear the target immediately after their attempt.
Multi-tenant institution and enrolment management, content publishing, role-based access, bulk import and scheduled export. The unglamorous infrastructure that makes the rest deployable in a real school.
Engineering practice
Schools cannot run on a prototype. Everything we ship is designed for the day it is someone else's responsibility to keep running.
Happy to go deeper on architecture, evaluation methodology or data handling for grant, partnership or procurement review.