AIPLA

An AI-Native Learning Platform for Danish Physics Education

The University of Copenhagen's Department of Science Education needed a generative-AI platform for upper-secondary physics teaching that would survive a data-protection review, skeptical teachers, and a shared-phone classroom — not a chatbot bolted onto a curriculum. Sunholo leads the platform engineering, with the first classroom activity live in week one and weekly releases since.

Education GenAI Platform GDPR by Design

AIPLA chat and workbench interface showing a Socratic physics tutor paired with an interactive projectile-motion simulator

The first live release, May 2026 — the platform has shipped weekly since

4-Month Build First Activity Live in Week One

Weekly Releases Continuous Iteration, Not a Single Launch

No Student Accounts Group-Based Access, EU-Hosted

Cloud → Institution-Owned Every Service Has a Transition Path

The Partner: University of Copenhagen

AIPLA is a three-year research project hosted by the Department of Science Education at the University of Copenhagen, investigating how generative AI can support teaching, learning, and assessment in upper-secondary physics.

The project's own framing is deliberately restrained: "the aim is not to maximise the amount of AI used in a lesson." It works with Danish physics teachers to design and study activities where AI supports students "without replacing the reasoning, experimentation, and communication through which physics is learned." Sunholo leads the platform engineering — turning that design principle into a working platform that real teachers use with real classes, not a research prototype that stays in a lab.

The Challenge

Real Pedagogy, Zero Guardrails

Danish physics teachers had already discovered generative AI works for building teaching material. One teacher's own trials — a Socratic tutor paired with a hand-built physics simulation — turned out to be exactly the form factor the platform needed to support at class scale. But none of it was safe to put in front of a room of 16-year-olds.

No log capture — chat ran directly browser-to-provider, invisible to teachers and researchers
Browser-exposed API keys — one leaked key, one unbounded bill
Every teacher reinventing the same simulation-plus-tutor pattern alone
No path through data-protection review — a real deployment had to survive GDPR scrutiny before a single student touched it

The bar was clear: build the infrastructure that makes teachers' own best ideas safe, configurable, and reusable — without losing what made them work.

AIPLA workbench showing an interactive projectile-motion simulator with velocity, angle and gravity controls

The workbench surface from the first release (May 2026) — a live simulator the tutor reads and responds to in real time

The Sunholo Solution

Built on the AI Protocol Platform

AIPLA runs on Sunholo's own open-source AI Protocol Platform template — a maintained foundation, not a one-off build. The template's own promise: "clone, set LOCAL_MODE=1, run make dev, chat in under 30 minutes with zero GCP credentials."

Activity-First, Not Chatbot-First

The platform's core design principle: "the core product is an activity, not a chatbot — a learning goal, teacher-prepared context, tutor behaviour, and a connected workbench designed together." The AI directs where each output lands — chat for conversation, workbench for the live simulator — reading the workbench state in real time.

Content-Review Pipeline

Every AI-generated simulation passes a server-side safety pipeline — tag allow-lists, static analysis for outbound network calls, a headless render check — before it reaches a sandboxed iframe with no same-origin access, no top navigation, and a locked-down content-security policy.

Central Budget & Key Management

Students and teachers never see an API key — "central AI service configuration rather than browser-entered keys." Every model call routes through project-managed keys, with per-group and per-class budget caps enforced at the router so one class can't drain another's allocation.

Model Routing, No Lock-In

Model choice varies by environment and task, weighed on "task capability, modality, operational constraints, data requirements, latency, and cost" — cloud and locally-hosted models are interchangeable through configuration, not a fixed default baked into the app.

Built for Institutional Handover

The platform is designed for service replacement from day one: teacher identity, model inference, document retrieval, analytics, storage, logging, and deployment all have a documented transition path onto university-owned infrastructure, gated on institutional readiness rather than left as a someday problem.

Group-Based, Not Individual Accounts

Students join with a group code rather than a personal account — data minimisation built into the identity model, not bolted on for compliance

Versioned, Reviewable Artefacts

Interactive activities are versioned so they can be inspected and moved — teachers review and approve before students see them

Separate Dev, Test & Production

Changes move through independently-tested environments in EU regions with automated checks, not straight from a laptop to a live classroom

More Than a Chat Box

Activity formats span simulation, virtual lab, conceptual dialogue, document and image submission, tables, calculators, and concept maps

Measurable Impact

Week 1 → Month 5

Built in the Open, Iteration by Iteration

The first complete classroom activity was live in week one. By week six the platform had grown from three hand-built lessons into a full authoring and sharing environment teachers could build their own activities in — with session sync and classroom operations shipping in the months since. No single big-bang launch — a visible build log the whole way.

3 Physics Activities Live

Projectile motion, Planck's constant, and kinematics — each pairing a simulation with a Socratic tutor

Teachers Author, Not Just Choose

What began as three fixed examples became a platform where teachers configure their own activities

No Student Accounts

Group-based access designed to hold up to institutional data-protection review, not bolted on after the fact

Open Foundation

Built on Sunholo's public AI Protocol Platform template, not a closed one-off codebase

Technology Stack

Platform

Multivac AI Platform AI Protocol Platform (open source) Google Cloud (EU regions)

AI & Models

Multi-provider routing Chosen per task by capability, cost & latency

Integration

MCP & open interaction protocols Versioned interactive artefacts Separate dev / test / production

Building an AI Platform That Has to Survive Scrutiny?

If your AI deployment needs to hold up to compliance review, a skeptical audience, and real-world usage beyond a demo, Sunholo can help you build it right the first time.

Schedule a Consultation View More Case Studies

An AI-Native Learning Platform for Danish Physics Education

The Partner: University of Copenhagen

Building an AI Platform That Has to Hold Up?

The Challenge

Real Pedagogy, Zero Guardrails

The Sunholo Solution

Built on the AI Protocol Platform

Activity-First, Not Chatbot-First

Content-Review Pipeline

Central Budget & Key Management

Model Routing, No Lock-In

Built for Institutional Handover

Group-Based, Not Individual Accounts

Versioned, Reviewable Artefacts

Separate Dev, Test & Production

More Than a Chat Box

Measurable Impact

Built in the Open, Iteration by Iteration

Technology Stack

Platform

AI & Models

Integration

Further Reading

Case Study

Platform

Building an AI Platform That Has to Survive Scrutiny?