A hardware-software stack for AIED applications

Overview

Artificial Intelligence in Education is moving beyond isolated AI tools toward systems that can support teachers and students directly in real learning environments. Guerino et al., with Seiji Isotani, describe this through Augmented Intelligence: AI should support human decision-making rather than replace it, while real deployments also have to account for reliability, teacher involvement, and access to technological resources.

EdgeAIED explores the infrastructure side of that problem. Instead of treating every AIED application as a separate cloud service, the project investigates a reusable hardware-software stack that can run educational AI applications locally on edge hardware. The goal is to provide a common platform for vision, audio, language, and feedback pipelines while reducing dependence on external computers, persistent connectivity, and application-specific hardware.

The broader idea is to make AIED systems easier to prototype, deploy, and reuse in classrooms, especially where cost, connectivity, privacy, or infrastructure make cloud-first systems difficult to scale.

Research question

How can we run useful AIED applications on local, offline and low-cost hardware in resource constrained classrooms?

Previous work

A lot of EdgeAIED builds on Seiji Isotani’s AIED Unplugged work. The main idea is simple: educational AI should still work in classrooms with weak internet, limited devices, and limited compute, instead of assuming everything can run in the cloud.

Later work from Uema, Isotani and collaborators pushed this further with the idea of pedagogical elasticity: the system should adapt what it can do based on the hardware available. Their offline-first work also shows that smaller local models can be more practical on constrained devices.

Other papers from the same research group show the limits more clearly. Santos et al. looked at running LLMs on resource-constrained smartphones, while Souza et al. explored handwritten-equation detection on disconnected low-cost devices. MathAIDE also looked at how teachers viewed an unplugged intelligent tutoring system for underserved regions.

EdgeAIED takes these ideas and tries to make them reusable in one place: a single local hardware platform where different models, runtimes, and AIED applications can be tested without changing the learner-facing device.

Pipeline

The pipeline is designed so the whole interaction can stay on the device. A student can enter a prompt through the touchscreen or microphone. Voice is transcribed locally with Whisper, then passed into the EdgeAIED prompt manager, which adds the application logic and any relevant local context before sending it to the on-device Qwen model.

The response then comes back through the same local stack. It can be shown directly on the display or spoken through Piper text-to-speech and the built-in speaker. Underneath this interaction are the application, inference, and system layers that handle the EdgeAIED interface, model runtime and quantization, Linux drivers, audio, storage, and hardware access.

EdgeAIED local multimodal software pipeline

Hardware Diagram

EdgeAIED brings together local compute, a touchscreen, microphone, speaker, wireless connectivity, and general-purpose I/O into one self-contained device. The Radxa CM5 handles local inference and the Linux software stack, while the surrounding hardware provides the audio, display, power, and connectivity needed to run classroom AIED applications without relying on an external computer.