Free AI for All is CUHK’s connected educational AI ecosystem. With one shared identity, students, teachers and teaching assistants move from multimodal chat, to AI that uses tools, to supervised agents, and finally to human-reviewed assessment support. The design pairs free access with practical AI literacy: asking better questions, checking outputs, managing cost, supervising actions and retaining responsibility for the result.

Evidence note. These images were captured from authenticated demo environments on 14 July 2026. Credentials and real student data are not shown. The GradePilot captures use a demo submission; no score or feedback field was changed or saved. GradePilot is an early-stage prototype, and no pending submission is presented as a completed or AI-approved grade.

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 original contribution is the educational design around current AI systems: equal access, a clear progression of capability, role-specific learning, cost control, practical guidance and continuing human review.

1. ChatNextWeb: a simple route into multimodal AI

The first stage gives users an approachable conversation surface before they progress to more complex tools. The captured profile is the teacher-facing demo. Its interface brings together conversation history, reusable prompts, image input, model controls and a familiar chat box.

Authenticated CUHK SCI Chat interface with conversation history, prompt controls and a large message field.
Figure 2. Guided multimodal chat. A low-friction interface lets educators focus on asking, inspecting and improving prompts instead of configuring separate commercial accounts.

Model choice is visible rather than hidden. Users can compare systems from several providers and discuss why different models give different answers, where each may fail, and when a lower-cost model is sufficient.

CUHK SCI Chat model selector showing multiple OpenRouter, OpenAI, Poe and Anthropic models.
Figure 3. Models become an object of study. The selector supports deliberate comparison across providers and capabilities instead of treating AI as one opaque answer engine.

2. OpenWebUI: AI that can use tools

The second stage moves beyond conversation. The authenticated OpenWebUI prompt bar exposes native web search, image generation, reasoning, voice input and file or extension controls. The selected-model description also surfaces price and data-handling information, reinforcing that capability, privacy and cost should be considered together.

Authenticated CUHK OpenWebUI screen showing a selected model, model price and privacy description, and web search, image generation and reasoning controls.
Figure 4. Tools are available at the point of use. Search, image and reasoning controls make it possible to teach not only what AI says, but how a user chooses, checks and combines its tools.

The searchable catalogue groups models by provider and service tier. This makes model selection explicit and supports the project’s customised cost metering and weekly budget controls rather than defaulting every task to the most expensive option.

OpenWebUI model catalogue with provider and service-tier filters and a list of available AI models.
Figure 5. Choice with cost awareness. Users can select an appropriate model while the institution manages access and expenditure centrally.

OpenWebUI also renders technical notation directly in a teaching conversation. In this short statistical-mechanics example, the user asked for two display equations and a compact explanation; the result preserves the mathematical structure instead of returning source markup or a code block.

Authenticated OpenWebUI conversation rendering the partition function and mean energy of a two-level system as two display LaTeX equations.
Figure 6. LaTeX rendered as readable mathematics. The response turns a precise natural-language instruction into two displayed equations, including the inverse-temperature definition and the differentiated partition function.

A second prompt demonstrates multiple tools inside one response. Native web search finds the exact SI values of Planck’s constant and the speed of light from official sources; the Code Interpreter then multiplies them and returns the result in a structured table. The visible tool trace, source domains, calculation and final values allow a learner to inspect the path as well as the answer.

Authenticated OpenWebUI response showing a multi-tool turn with web research from NIST and BIPM, Code Interpreter execution, equations and a three-row results table.
Figure 7. Web search and computation in one turn. Two official-source lookups are combined with an executed calculation of hc, making source checking and tool orchestration visible in the same response.

Image work is similarly iterative. The first prompt generated a clean educational illustration of a double-slit experiment. The result was then supplied to the image-edit workflow with explicit preservation and change instructions: retain the apparatus and perspective, move to a white background, change the beam to magenta, keep the gold fringes and add a red directional arrow.

AI-generated isometric double-slit experiment with a dark navy background, teal beam and gold interference fringes.
Figure 8. Initial educational image generation. A text prompt produces a coherent, text-free scientific illustration with specified colours, apparatus and interference pattern.
Edited double-slit illustration with a white background, magenta beams, retained gold fringes and a red arrow pointing toward the pattern.
Figure 9. Instruction-guided image editing. The edited version visibly changes the palette and adds the requested arrow while preserving the experiment's subject and isometric visual language.

The working authenticated image capture uses OpenWebUI’s available google/gemini-3.1-flash-image alias. The requested preview-labelled OpenRouter route, google/gemini-3.1-flash-image-preview, was also tested twice in ChatNextWeb, but its current adapter returned a JSON parsing error; it is therefore not represented here as a successful run.

3. Claude Code and Codex: supervised agents in a working environment

The third stage places agents inside an authenticated JupyterLab environment. Users can work with files, notebooks and terminals while keeping the agent’s actions visible. This turns AI literacy from answer checking into process supervision: reviewing plans, tool calls and changes before accepting them.

JupyterLab terminal showing the Codex command-line interface and commands for review, sandboxing, resuming and applying changes.
Figure 10. Codex inside JupyterLab. The command surface exposes review, sandbox, resume and apply workflows, providing clear points for human inspection during multi-step work.

The same environment also provides Claude Code. Its interface exposes agent, tool and permission controls; the screenshot is a capability check, not an autonomous task. Students and educators can therefore compare agent behaviour within one supervised workspace.

JupyterLab terminal showing Claude Code help with agent, tool and permission controls.
Figure 11. Claude Code in the same supervised workspace. Multiple agents can be taught and evaluated through a common environment while the user remains responsible for what is allowed and accepted.

4. GradePilot: early-stage, human-reviewed assessment support

GradePilot connects course materials, assignments, submission and review. The student-facing demo shows structured homework cards, problem access, due dates and a clear submission route. This creates the workflow into which rubric-guided draft feedback can be introduced without bypassing the teaching assistant.

GradePilot student assignment dashboard showing three homework cards, problem and submit buttons, points and due dates.
Figure 12. Assessment begins with a clear human workflow. Students see what is due and submit through one course interface; the system is not presented as an answer generator.

Each submission keeps a visible status and version history. The demo item below is still pending, which accurately reflects the prototype stage. GradePilot may draft rubric-guided feedback and a provisional grade, but a teaching assistant must inspect, edit or reject that draft and retains every final decision.

GradePilot submission details dialog showing submitted status, PDF preview, download controls and three recorded versions with score fields.
Figure 13. Traceable review, not automatic authority. Status, source file and version history remain visible so draft assistance can be checked against the student's actual submission.

The TA-authenticated grading screen makes final responsibility concrete. It presents an overall score, sub-question rubric fields, written feedback, the submitted source and a rendered-PDF pane together. “AI Fill” is an optional draft action beside a separate “Save” control; the teaching assistant remains in the review loop and decides what, if anything, becomes part of the grade.

Authenticated GradePilot teaching-assistant grading screen for a demo submission, showing overall and per-question score fields, feedback, source and rendered-PDF panes, AI Fill and Save controls.
Figure 14. The grade screen keeps the human decision visible. The rubric, evidence and draft-assistance control share one review surface. This capture opened the demo record read-only: no AI Fill, score entry or save action was performed.

What the evidence demonstrates

Award area Evidence on this page
Innovation A coherent progression from multimodal chat to tool use, supervised agents and human-reviewed assessment, joined by shared access and practical AI literacy. The project does not claim to have built the underlying foundation models.
Impact and potential impact The interfaces reduce entry barriers while supporting distinct student, teacher and TA workflows. Application records report more than 300 active users across at least ten Physics and Statistics courses, approximately 100 daily users, more than 40 billion tokens served and over 100 written feedback responses.
Responsible depth Visible model choice, price and data-handling information, institutional metering, inspectable tool traces, agent sandbox and tool controls, versioned submissions and retained TA authority create repeated opportunities to question, verify and intervene.

The 2024-25 survey records reported that more than 75% of respondents found the service easy to use and more than 80% reported improved learning. Those figures are reported outcomes; the screenshots above demonstrate the actual learning and review surfaces through which the programme operates.