Evidence hub · Innovation · Impact · Depth

The innovation in one view

Free AI for All is one connected educational ecosystem rather than a collection of unrelated tools. The same university identity opens a progression from multimodal conversation, to AI that uses tools, to supervised agents working inside a computer environment, and finally to early-stage assessment support in which a teaching assistant retains every final decision.

Diagram showing four connected stages: ChatNextWeb multimodal chat, OpenWebUI tool-using AI, supervised agents in Claude Code and Codex, and GradePilot human-reviewed assessment.
Figure 1. One identity, four connected stages. The ecosystem connects equal access, a progression of current capabilities, role-specific learning, institutional cost control and continuing human review.

The project does not claim to have created the underlying foundation models. Its original contribution is the educational and technical design around them: free shared access, a coherent capability progression, distinct student, teacher and TA workflows, practical AI literacy, per-user cost metering and an evidence loop that informs the next iteration. This design was necessary because access alone does not teach people how to question outputs, supervise actions or remain accountable for decisions.

1. ChatNextWeb: multimodal AI made approachable

ChatNextWeb provides a familiar entry point for text-and-image conversations. Model choice is visible rather than hidden, allowing students and educators to compare systems, investigate why answers differ and decide when a lower-cost model is sufficient.

CUHK SCI Chat model selector showing multiple OpenRouter, OpenAI, Poe and Anthropic models.
Figure 2. Models become an object of study. The selector supports deliberate comparison across providers and capabilities instead of presenting AI as one opaque answer engine.

2. OpenWebUI: AI that can use tools

OpenWebUI moves beyond conversation to document retrieval, web-assisted work, image generation and editing, reasoning and in-browser code execution. Its searchable catalogue makes model selection explicit. The customised service also meters each student’s spend against a weekly budget cap, joining equitable access with cost-aware use.

OpenWebUI model catalogue with provider and service-tier filters and a list of available AI models.
Figure 3. Choice with cost awareness. Users can select an appropriate model while the institution manages access and expenditure centrally.

The working environment keeps tool traces and results inspectable. In the example below, native web search finds exact SI values from official sources and Code Interpreter performs the calculation; the source domains, executed step and result remain visible together.

Authenticated OpenWebUI response showing web research from NIST and BIPM, Code Interpreter execution, equations and a results table.
Figure 4. Research and computation in one visible workflow. Learners can inspect both the answer and the path that produced it.

Image generation is also treated as an iterative, reviewable process. The first prompt creates an educational double-slit illustration; the second preserves the apparatus while changing the background, beam colour and annotation.

English OpenWebUI chat showing an image-generation prompt and an AI-generated isometric double-slit experiment.
Figure 5. Educational image generation in context. The prompt, selected model, usage meter and result are visible in the same conversation.
English OpenWebUI chat showing an image-edit instruction and the edited double-slit illustration.
Figure 6. Instruction-guided image editing. The revised image can be compared directly with the preservation and change instructions.

3. Claude Code and Codex: supervised agents

Claude Code and Codex place agents inside an authenticated JupyterLab workspace. Users can inspect files, plans, commands, tests and artifacts while keeping control of what the agent may change. This extends AI literacy from checking an answer to supervising a multi-step process.

English JupyterLab terminal showing Codex completing a three-point review plan for an Euler-method harmonic oscillator notebook without editing or running files.
Figure 7. Codex completes a bounded review task. The prompt explicitly limits the task to analysis, and the model, request and completed response remain visible together.

Codex supervised file-and-test demonstration. Codex inspects a plotting script, creates two tests, runs them with visible approval, regenerates a PDF and opens the result from the JupyterLab file browser.

English JupyterLab terminal showing Claude Code completing a three-point review plan for an Euler-method harmonic oscillator notebook without editing or running files.
Figure 8. A comparable Claude Code task. The working environment lets educators compare agent behaviour using closely matched prompts.

4. GradePilot: assistance without automatic authority

GradePilot connects assignments, submission history, grading and feedback. It is a working early-stage prototype under evaluation. AI Fill is an optional action that can draft rubric-guided scores and feedback, while a separate Save control leaves the teaching assistant responsible for inspecting, editing or rejecting the draft and making every final decision.

Authenticated GradePilot teaching-assistant grading screen for a demo submission, showing score fields, feedback, source and rendered-PDF panes, AI Fill and Save controls.
Figure 9. Human review is built into the interface. Rubric fields, the student's evidence, draft assistance and the final save action share one review surface. No AI Fill, score entry or save action was performed for this capture.

How this differs from existing approaches

Most commercial AI products are designed for one individual and one interaction type. This project is designed for an educational community: students, teachers and TAs use the same connected ecosystem but have different responsibilities and learning goals. Free access and cost control address economic inequity; role-specific tutorials and visible tool traces address capability inequity; supervised agent actions and retained TA authority address accountability. The system can adopt new models and interaction modes without abandoning those educational safeguards.

Evidence hub · Innovation · Impact · Depth