The intelligence layer behind every score.
A real AI Evaluation Engine and AI-assisted roadmap pacing are live in the platform today — grounded in each school's own curriculum, with a fallback that never fakes a result.
Two AI systems, live in production.
AI Evaluation Engine
A single OpenAI gpt-4o-mini call scores every free-text answer across many real dimensions — correctness, reasoning, terminology, confidence, misconceptions, depth, completeness, and Bloom level — in one pass. Delivered via OpenAI's Batch API so AI grading stays affordable at real scale, not a synchronous call per answer.
AI-Assisted Roadmap Pacing
Sequences a chapter's concepts into a real day-by-day plan against the institute's own academic calendar. Every AI-generated plan is validated hard — every concept placed exactly once, nothing hallucinated — and falls back to a deterministic pacer on any validation failure, so a roadmap request always gets a complete, usable answer.
Evaluation, scored on eleven dimensions.
Every answer is assessed far beyond right/wrong — including reasoning, terminology, and misconceptions, all in one real evaluation call.
Every question tagged by Bloom's level.
This lets a profile express nuance a percentage never could: excellent memory, poor application, good analysis, weak creativity.
What the AI layer grows into.
AI Tutor
RoadmapAnswers strictly from the school's own curriculum, grounded via retrieval-augmented generation — not the open internet.
Predictive Analytics
RoadmapForecasts exam score, dropout risk, and knowledge decay before they become a problem.
Recommendation Engine
RoadmapPlain-language next steps for teachers, students, and parents, generated automatically from the profile.
Teacher Assistant
RoadmapSummarizes class-level insight for teachers — patterns across a whole roster, not one student at a time.
Parent Insight Generator
RoadmapTranslates raw scores into plain-language updates for parents.
How it's built.
Large Language Model — running today
OpenAI, direct API — synchronous for AI-assisted roadmap pacing, and Batch API for assessment evaluation, so grading stays affordable at real scale.
Embeddings, RAG & predictive ML — planned
An embedding model and retrieval-augmented generation to ground the AI Tutor, and dedicated ML models for predictive analytics — exam score, dropout risk, retention decay.
Grounded, not generic — and never a fake result.
Every evaluation is grounded in the school's own curriculum, not open-internet knowledge. And if a response genuinely can't be scored, it stays visibly ungraded — the same fallback-safe design behind AI-assisted roadmap pacing, which never fakes a plan either. Nothing in the AI layer papers over a failure with a default value.