AI Tutoring Platform Development

LLM-powered subject-specific Q&A scoped to your curriculum -- the tutor stays on topic and references your content, not the open internet. We use retrieval-augmented generation (RAG) over your ingested course material so responses are grounded in your actual syllabus, not the open internet. Socratic questioning mode is implemented via structured prompt chains using GPT-4o or Claude 3.5 Sonnet: instead of handing over an answer, the tutor generates a sequence of leading questions calibrated to the student's last response, nudging them toward the reasoning rather than the result. Configurable persona and tone by age group -- a primary school science tutor sounds different from a professional certification prep tutor. Session memory keeps the conversation coherent across a single study session using a windowed message history with summarisation to stay within context limits. Learning objectives from Bloom's taxonomy are mapped to each curriculum topic (remember, understand, apply, analyse, evaluate, create), so the tutor adjusts question depth to match the intended cognitive level rather than asking recall questions when application is the goal.

Performance-based difficulty adjustment that moves students forward when they demonstrate mastery and slows down when they don't, rather than advancing everyone on a fixed schedule. The underlying knowledge tracing uses Bayesian Knowledge Tracing (BKT) for binary skill mastery estimation or Deep Knowledge Tracing (DKT) with LSTM networks for more granular, multi-concept mastery modelling across sequential exercises. Prerequisite mapping built on your curriculum graph ensures gaps in foundational concepts are filled before a student is pushed into more advanced material -- you cannot be routed to quadratic equations without demonstrating proficiency in linear algebra first. Branching content sequences adapt based on assessment results so two students in the same course can follow genuinely different paths through the same material. Adaptive difficulty adjustment operates at item level: the system selects the next question from the content pool to sit just above the student's current estimated mastery -- neither trivial nor frustrating -- using item difficulty parameters from Item Response Theory (IRT). Spaced repetition scheduling is woven into the path for memorisation-heavy subjects, resurfacing weaker concepts at the optimal interval to prevent decay without over-reviewing mastered material.

Written response scoring with rubric alignment gives students specific, actionable feedback rather than a raw score. Code submission feedback runs student code against a hidden test case suite, surfaces which cases passed and failed, highlights where the logic breaks down, and offers targeted hints generated by an LLM prompt chain -- without giving the answer away. Essay feedback addresses argument structure, evidence use, and coherence against your assessment criteria, comparing submissions against exemplar answers using embedding similarity rather than keyword matching to catch paraphrased but correct reasoning. Misconception detection is built on Item Response Theory (IRT): by tracking which specific distractors students select on multiple-choice items, the system identifies systematic misunderstandings -- not just gaps -- and generates corrective explanations targeted at the specific misconception pattern. Feedback is graded in specificity based on student tier: a learner struggling with fundamentals gets more scaffolded guidance, while a high-mastery student gets a challenge question instead. All feedback is logged against the xAPI learning record store (LRS) so teachers can audit the feedback each student received, not just their final scores.

Per-student and per-class analytics that show which learning objectives have low mastery across your cohort -- not just which students scored low on the last test. The mastery model draws from the BKT or DKT knowledge tracing layer, giving teachers a probabilistic estimate of each student's mastery per concept rather than a raw percentage score. Teacher dashboards surface actionable gap reports showing the five lowest-mastered objectives in the cohort, with drill-down to the individual student level, so instructors know where to focus live teaching time rather than guessing from end-of-unit tests. Early warning alerts use configurable mastery thresholds: if a student's estimated mastery on a prerequisite concept drops below a set level with N days until the related assessment, the teacher is notified by email or in-app notification, giving time to intervene before the summative evaluation. All progress data is emitted as xAPI statements to your learning record store (LRS), which means full interoperability with any LRS-compatible analytics or reporting tool -- Watershed, Learning Locker, SCORM Cloud, or your own data warehouse. WCAG 2.1 AA compliance is applied to all teacher-facing dashboard views, including screen reader support for chart annotations and keyboard-navigable data tables, so the analytics tools are accessible to instructors using assistive technology.

Vocabulary and concept decks with SM-2 algorithm scheduling, calculating the optimal next review interval per card per student based on recall accuracy and ease factor. Both student-generated and teacher-generated decks are supported, with teacher decks scoped to specific learning objectives from the curriculum map. Performance-based card scheduling adjusts the ease factor and interval for each card individually: a card answered correctly multiple times in a row is pushed further into the future, while a card that trips a student up is rescheduled within 24 hours regardless of its previous interval. The SM-2 ease factor adapts to each student's recall pattern -- a student who consistently struggles with irregular verb conjugations will see those cards resurface at shorter intervals than a peer who finds them easy. Progress is tracked by deck, subject, and learning objective so the spaced repetition data feeds back into the mastery model, giving the knowledge tracing layer a signal from review sessions as well as from formal assessments. All review activity is emitted as xAPI statements to the LRS, which means the spaced repetition data joins the same longitudinal learning record as quiz results, tutor conversations, and submission feedback -- giving teachers one coherent view of each student's progress rather than siloed tool data.

Connects to your existing LMS -- Moodle, Canvas, Blackboard, or custom-built -- via LTI 1.3 so the AI tutor launches inside your existing learning environment without requiring a separate login. LTI 1.3 uses the Assignment and Grade Services (AGS) extension to write scores back to the LMS gradebook automatically, so teacher reporting is not fragmented across two systems. Tutor interactions and assessment results are mapped to your learning objectives and emitted as xAPI statements to your learning record store (LRS), giving you a persistent activity record that survives a future LMS migration. The xAPI LRS also makes it straightforward to combine tutor data with data from other tools -- video lesson completions from a Wistia integration, quiz results from your assessment engine, live session attendance -- into a single longitudinal learner record. SCORM-compatible content rendering is included where your existing content requires it, and SCORM completion events are normalised into the same xAPI schema as native tutor activity. Personalized learning path recommendations are generated from the LRS activity history: the path engine reads the learner's full xAPI event log, identifies the lowest-mastery objectives, and recommends the next content sequence accordingly -- making the learning path responsive to what a student has actually done across every tool, not just their scores in the tutor.