Methodology document directory

Aida Learning Companion

Agentic AI learning companion, guiding students through the four stages of the AIDA framework (Ask → Imagine → Design → Act) to explore independently. Integrates Bloom's cognitive tracking, course material retrieval and learning alerts.

AIDA frameworkAgentic AISocratic guidanceLearning alerts

Deep Research

For complex learning problems, a ReAct multi-step research mode (Plan → Action → Observation → Reflect → Synthesize) that gathers evidence across tools and integrates it into cited answers. Includes five layers of cost and safety safeguards, applicability checks, and quota and billing mechanisms.

ReActMulti-step loopAuto-judgeFive-layer defenceCitation assembly

Bloom's Taxonomy cognitive level analysis

Automatically classify the cognitive level of Student messages through LLM (memorisation→understanding→application→analysis→evaluation→creation), including meaningfulness judgements, dialogue context handling and multimodal support.

LLM classificationReal-time + batchPrompt version management

Knowledge Graph Analytics

Extract knowledge concepts and relationships from teaching materials, and analyse knowledge mastery from Student conversations (mentioned→exploring→understanding→applying→mastered) to track learning progress.

Concept extractionMastery trackingGraphRAG

Cross-course learner profile

Integrates 9 data sources (course records, conversations, Bloom's, personality scales, achievement, etc.); the LLM synthesises a structured learner profile and injects it into AI conversations to enable personalisation.

Multi-source integrationLLM synthesis24h cache

The educational implications of Uedu Science

Design rationale for the 96 browser-based scientific simulations: how inaccessible concepts are turned into operational models, how variable control preserves process traces, the value of juxtaposing real data with idealised models, and the work that the AI TA should not take over.

96 simulationsInquiry and practiceModel-based learningTA boundaries

RAG Teaching Material Retrieval-Augmented

Teacher uploads course materials → chunking (800 tokens) → vectorisation (text-embedding-3-small, 1536 dimensions) → cosine similarity retrieval → inject into AI responses.

EmbeddingSemantic chunkingTop-K retrieval

Semantic vector

Calculates a 1536-dimensional word embedding for each Student message, supporting semantic similarity search, similar-question detection and research clustering analysis. Real-time + batch dual-path processing.

Cosine SimilarityScope cacheBatch backfill

Uedu Open

An AI conversational learning interface for OpenCourseWare (OCW), importing materials such as MIT OpenCourseWare, with prompt augmentation based on transcripts to upgrade static materials into an interactive learning experience.

OCW integrationGPT-5.4-miniTranscript RAGSSE streaming

Research field map

Topic tag cloud, teacher public profiles, external paper search via OpenAlex, and research clusters generated by semantic grouping.

Research clusterOpenAlexInstructor profile

School version

A university-level student community forum, with participant identity scoped by university_code. Supports anonymous dual mode (fixed alias within threads), display-name snapshots, staggered rename cooldowns, 16 emoji reactions, and moderator review.

Anonymous dual modeSnapshot attributionEdit historyModerator review

Collaborative note-taking

A multi-user real-time collaborative document based on Yjs CRDT (backend pycrdt), with Socket.IO incremental updates, support for snapshot-based version history, line-based comments, clipboard image uploads, and export to the Course RAG knowledge base.

Yjs CRDTSocket.IOVersion snapshotExport to RAG

Socratic Dialogue pre- and post-test

AI generates questions automatically, pairs pre- and post-tests, and uses Normalised Gain to estimate learning change.

Pre/post test pairingNormalised GainAI question generation

Simulated Debate Attitude Survey

Likert five-point scale, covering three dimensions: position, confidence, and openness, with reverse-item processing and attitude-change reporting.

Likert 5-pointThree dimensionsReverse-coded items

Competency-based Worksheet

From classroom audio transcripts, RAG teaching materials or instructor text, AI automatically generates five-level Worksheets aligned with the 108 Curriculum's competency-based approach (context introduction→reading comprehension→exploration→reflection→transfer). Mixed marking: multiple-choice questions are marked instantly; open-ended questions are pre-marked by AI and then confirmed by the instructor.

108 CurriculumFive-level literacyHybrid markingPDF export

AI interactive assignment

The Instructor sets the interaction mode and assessment criteria; Students complete the assignment in conversation with AI; LLM automatically grades according to scoring_criteria; the Instructor can review and revise.

Custom marking rubricLLM auto-markingInstructor review

AI Quiz Generator

Automatically generate questions based on the teaching content, supporting multiple question types and difficulty distributions (easy/medium/hard), and call the LLM in batches by “question type × difficulty”.

Difficulty distributionBatch generationMultiple question types

Forum Scoring

Weighted scoring system (posts/comments/replies), emoji bonuses (thumbsup/confetti/rocket) and penalties (zzz), instructor review mechanism (eyes), BERT quality evaluation.

Weighted scoringEmoji mechanismBERT evaluation

Prompt Leaderboard

An AI prompt blind-comparison system similar to LM Arena, using ELO ratings (K=32, initial 1500), with at least 30 matches required before inclusion in the ranking.

ELO scoringBlind A/B testing5 question types

Learning Profiling

Three open-source psychology scales: Holland RIASEC (48 items), IPIP Big Five (50 items, including reverse scoring), OEJTS/MBTI (32 items), with full scoring algorithm notes.

RIASECBig FiveOEJTS

UCG cognomics quiz toolkit

14 scientifically validated cognitive tests, covering 6 constructs (executive function, processing speed and attention, working memory, spatial cognition, verbal ability, metacognition), measure learners' cognitive ability traits in a web-based way.

CognomicsExecutive functionWorking memory14 Instruments

PALM Physiologically-Aware Language Models

Integrates Garmin wearable data (sleep, HRV, stress, Body Battery); dual-baseline z-score deviation detection (7-day acute + 30-day trend), injected into AI conversations.

Garmin ETLDual baselineZ-Score

UeduPAD user guide

Build a research-grade iOS app based on the Garmin Health Companion SDK: install and sign in, pair the watch, stream beat-by-beat physiological signals in real time (BBI / heart rate / respiration rate / stress), record and upload in intervals, multi-participant measurement workflow, and research-ethics governance.

iOS data collection toolGarmin SDKBBI streamingUser manual

Data export and data governance

Three electronically signed consent forms (teacher responsibility form, export consent form, Super TA NDA), encrypted export process, notification under Personal Data Protection Act §41, IRB compliance framework.

Informed consentEncrypted exportPersonal Data Protection ActIRB

School Management Console User Guide

For use by each school's campus administrators (campus_admin) and campus viewers (campus_viewer): access and permissions, usage statistics dashboard summarised by semester, campus-level API Key (BYOK) settings, administrator assignment, and K-12 roster-based course creation.

Administrative managementBYOKUsage statisticsUser manual

Uedu Labs laboratories and research projects

The governance unit of the research team: roles and permissions (PI/members), processes for joining and establishing a team, the participant-led authorisation mechanism for research projects (mirror view, access records, withdrawal at any time), and the boundary between service use and research use.

Research governancePI modelProject authorisationMirror transparency

MCP Server

Package Public API v1 with Model Context Protocol for LLM client use, supporting stdio and streamable-http dual transport, SHA-256 hashed keys, and two-tier rate limiting (per minute / per day), strictly separating public and sensitive data boundaries.

Anthropic MCPDual transportSHA-256 keyData boundary