Intro & scope
Use cases for teaching, learning, assessment and institutional processes in schools, universities and corporate training (EU/GDPR focus). Three patterns: EDU.CONTENT (generating material for teachers/trainers: 001, 006, 009), EDU.LEARN (learner-facing tutors, practice and study tools: 004, 010, 011, 016) and EDU.ASSESS (assessment, analytics and admin decisions: 002, 003, 007, 008, 012-015, 017).
Scope boundaries: 001 = course material from curricula, 006 = corporate/localised microlearning, 009 = teacher-side lesson planning and differentiation. 002 creates items, 003 grades work, 008 secures exams and detects misconduct. 004 = curriculum-grounded tutor, 016 = learner-owned study tools from own materials, 005 = path sequencing. 011 = free conversation practice, 012 = scored speech assessment. 014 = credit recognition of enrolled/prior learning, 015 = admissions screening.
AI Act: Annex III point 3 (access/admission, evaluating learning outcomes, steering learning, proctoring) makes 008, 012 and 015 high-risk; 003, 004, 005, 007, 013, 014, 017 become high-risk once outputs steer grades, placement or access decisions. Content generators are limited/minimal risk. Humans decide on every grade, integrity case, admission and recognition. Minors' data requires strict minimisation, no emotion inference at school or work (Art. 5).
Use cases
| ID | Title | Pattern | Maturity | Path | HW tier | AI Act risk |
|---|---|---|---|---|---|---|
| EDU-001 | Course & learning material generation | EDU.CONTENT | mainstream | A | HW2 | limited |
| EDU-002 | Quiz & exam question generation | EDU.ASSESS | mainstream | A | HW2 | depends |
| EDU-003 | Automated grading & formative feedback | EDU.ASSESS | early-adopter | A | HW3 | depends |
| EDU-004 | AI tutor & homework help | EDU.LEARN | early-adopter | A | HW3 | depends |
| EDU-005 | Adaptive learning paths | EDU.LEARN | mainstream | A | HW1 | depends |
| EDU-006 | Corporate training & microlearning content | EDU.CONTENT | mainstream | A | HW2 | limited |
| EDU-007 | Learning analytics & dropout early warning | EDU.ASSESS | early-adopter | A | HW1 | depends |
| EDU-008 | Exam proctoring & academic integrity | EDU.ASSESS | mainstream | A | HW2 | high |
| EDU-009 | Teacher assistant: lesson planning & differentiation | EDU.CONTENT | mainstream | A | HW2 | limited |
| EDU-010 | Simulation & role-play training (sales, medical, leadership) | EDU.LEARN | early-adopter | A | HW3 | depends |
| EDU-011 | Language learning conversation partner | EDU.LEARN | mainstream | A | HW2 | limited |
| EDU-012 | Oral reading fluency & spoken-language proficiency assessment | EDU.ASSESS | early-adopter | A | HW2 | high |
| EDU-013 | Classroom discourse analytics & AI teacher coaching | EDU.ASSESS | early-adopter | A | HW2 | depends |
| EDU-014 | Credit transfer, course equivalence & recognition of prior learning | EDU.ASSESS | early-adopter | B | HW3 | depends |
| EDU-015 | Admissions application screening & foreign grade conversion | EDU.ASSESS | early-adopter | B | HW3 | high |
| EDU-016 | AI study tools: notes, flashcards & spaced repetition from own materials | EDU.LEARN | mainstream | A | HW2 | minimal |
| EDU-017 | Early screening for dyslexia & specific learning difficulties | EDU.ASSESS | early-adopter | A | HW1 | depends |
Shared stack
Reference stacks per self-host HW tier. One platform serves the whole family and can be shared with KNOW (RAG), DOC (document parsing), SPEECH, LANG and CONV families.
| Tier | Reference stack | Serves |
|---|---|---|
| HW1 (CPU / edge / single small GPU) | scikit-learn/LightGBM, time-series features, Postgres + pgvector, on-device small ASR (whisper.cpp) and small LLM, LMS xAPI/Caliper event store | 005, 007, 017, on-device parts of 012 |
| HW2 (one 24-48 GB GPU) | vLLM or Ollama with an 8-30B instruct model, bge-m3 embeddings + reranker, pgvector/Qdrant, faster-whisper + pyannote, Piper/Kokoro TTS, Docling/OCR, Langfuse and guardrail layer | 001, 002, 006, 009, 011, 012, 013, 016, 008 |
| HW3 (multi-GPU / 70B+ and VLM) | vLLM serving a 70-120B reasoning model plus a VLM (e.g. Qwen-VL class), same embedding/RAG layer, streaming ASR/TTS or speech-to-speech for concurrent voice sessions, n8n/Temporal workflows with human-review queue | 003, 004, 010, 014, 015 |
Shared building blocks across tiers: - Content RAG: one curriculum/course-material index (Docling parsing, bge-m3, reranker) reused by 001, 002, 004, 009, 014, 016. - Learner model: one mastery/event store (xAPI to Postgres) feeding 005, 007, 004, 016 and 017; pseudonymised, role-based access, short retention. - Voice layer: one ASR/TTS/speech-to-speech service for 004, 010, 011, 012, 013, 016. - Human-review queue and audit log: mandatory for 003, 008, 014, 015 (grades, integrity, recognition, admission). - Guardrails and logging: age-appropriate filters, prompt-injection defence, AI-literacy notices (Art. 4, 50) and no emotion recognition. - Hosting: EU region or on-premises; DPIA template shared across the family.