University of Copenhagen · Center for Digital Education

AIPLA

AI in Physics Learning and Assessment

A three-year research programme investigating how generative AI can be meaningfully integrated into upper-secondary physics education in Denmark. This site collects the working plan for the four-month technical-infrastructure contract within it.

Pages

  • About — project context, the ADDIE method, the four research questions, intended outputs
  • Strands — the three work strands (pedagogical bots, simulation/games, scoping investigation) and how they map to the timeline
  • Examples — real artefacts a project teacher made with GenAI before AIPLA existed: a live interactive projectile sim, a Socratic system prompt, and AI-misconception illustrations
  • Timeline — 17-week plan, the three overlapping tracks, the handover fan-out, ownership map
  • Architecture — ADRs for Strand A: cloud provider, region, model, RAG store, routing layer, multimodal, student auth, AILANG disclosure
  • Evaluation — capability-floor framework: task taxonomy, model panel, KPIs that drive routing and the local-deployment migration timeline
  • Self-hosting — UCPH migration table: cloud components and their on-premises equivalents

People referenced by initials throughout (M, JB, AR, AD, DS, ZL, K, DK) for light anonymisation — the team knows the mapping. AD is the project postdoc, previously written as P2 before the hire.

Design principles

GDPR by design

EU-hosted infrastructure. No PII leakage to US providers where avoidable. Solutions hold up to UCPH data-protection review and to teachers’ and parents’ scrutiny.

Model-agnostic routing

Production chooses models per task class based on continuous capability-floor evaluation. Cloud APIs and local models are interchangeable through configuration.

Migration-ready

Every cloud component has a documented on-premises equivalent. Hosting can move to the University of Copenhagen’s own infrastructure when capability and policy align.

Handover as a first-class deliverable

Documentation, pairing, and ADR sign-offs from week 4, not week 13. Each major artefact has a named co-owner who will operate it after the contract.

Status

Post-freeze internal iteration (July). The mid-point review was held 26 June; the week-27 freeze has passed and the project is in the internal-iteration phase (Strand A iteration, Strand B build, Strand C scoping kickoff). Three student-facing physics skills are live — Boldkast (projectile motion, Danish stx), LED Planck (Planck’s constant, Danish stx), and KineBot (kinematics, NCERT/CBSE) — on a full teacher UI (classes, activities, analytics) with chat logs in BigQuery. The pre-freeze push added a teacher co-working co-pilot, an activity-authoring co-pilot, a rich-document workbench, image-based solution submission, cross-teacher activity sharing, and the physics sims served as portable MCP Apps. In early July the platform went to the partner-school teachers as a voluntary sneak peek — teacher login only, explicitly not for student use until the GDPR/consent work lands. A postdoc (AD) has been hired to lead the student perspective, officially starting ~1 October and already joining development meetings online. Teacher pilots begin mid-August, with a full-day teacher workshop at a partner gymnasium tentatively set for Friday 21 August; in the autumn the project group returns to the gymnasiums for observation, student interviews, and scoping the physical data-collection setup (cameras, audio). For the full shipped / in-progress / backlog breakdown — verified against code — see Current status on the Timeline page.