OUR STORY

From a student prototype to a platform for responsible AI play.

SpellTag asks a practical question: can generative AI make a physical spelling toy more responsive without replacing hands-on learning?

Start with the demo

THE OPPORTUNITY

Keep the value of physical play. Add a response that can change with every word.

Physical first

Children assemble a word with reusable alphabet cards instead of beginning with a blank screen.

Beyond fixed flashcards

A fixed picture library cannot respond to every idea. A controlled generative system can create a new visual interpretation.

Question the output

The image becomes a discussion object: what matched, what did not, and why might an AI interpret the word differently?

ZERO-COST INTERACTIVE DEMO

Try the AI Interpretation Lab

This experience uses examples generated previously. Revealing one does not make a new AI request.

1. Choose a stored example

Loading curated examples…

2. Build the word

CAT
In the physical prototype, the reader sends this word over Wi-Fi.

The learning happens after the image appears.

SpellTag does not present AI output as a guaranteed answer. It invites comparison, feedback, and discussion.

ONE CONNECTED LEARNING LOOP

From a physical card to a discussion about AI

The current prototype uses the existing ESP32-S3 and NFC reader, sending card events through the Wi-Fi data path with realtime delivery and a polling fallback.

  1. 1NFC alphabet cardPhysical word building
  2. 2SpellTag readerWi-Fi only
  3. 3Platform serviceEvent delivery + GPT Image
  4. 4Browser platformWord, gallery, missions
  5. 5AI illustrationCreative interpretation
  6. 6Learner feedbackQuestion, accept, or reject

NO-CALL TOY CONNECTOR

Step through one learning event

This simulator explains the connection only. It does not contact a device, database or image model.

Current step: Card tapped

  1. 1Card tapped
  2. 2Reader sends event
  3. 3Platform interprets
  4. 4Learner discusses
Technical path

PN532 reader → ESP32-S3 Wi-Fi station → device event delivery → Supabase Realtime with polling fallback → SvelteKit platform → GPT Image generation → gallery and feedback.

THE SOFTWARE PLATFORM

More than a one-off image generator

Two ways to participate

Use the Wi-Fi reader for the physical experience, or keyboard input when demonstrating the software platform.

Controlled visual generation

Production image generation is locked to GPT Image so provider choice and cost remain under operator control.

Learning missions

Word categories, missions, and progress cues turn isolated generations into repeatable spelling activities.

Feedback that improves the system

Learners can accept or reject a result and explain what went wrong, creating evidence for future prompt and product improvements.

RESPONSIBLE AI

Useful because the output can be questioned, not because it is always right.

Prompt and provider controls reduce risk, but no generative system can guarantee a safe or accurate result. SpellTag is designed for supervised use and makes disagreement part of the learning loop.

Read the safeguards and limitations

AI can misunderstand

An image may miss the intended meaning, especially for unusual or ambiguous words.

Interpret, do not memorise

The picture is a creative response, not a factual definition or trusted answer.

Controls plus feedback

Child-oriented instructions, provider restrictions and rejection reasons help manage risk.

Supervision expected

An adult should guide use, discuss surprising results and decide what is appropriate to keep.

PROJECT EVOLUTION

From course prototype to a stronger public platform

Stage 1

A+ final-year project

A complete NFC-to-image prototype established the core Spell → See interaction.

Stage 2

Open SPACE demonstration

Showcase feedback highlighted where input, dictionary quality, mobile access and the learning experience needed to mature.

Stage 3

Competition platform upgrade

Phrase input, PWA support and the Learning Journey are taking shape while dictionary, accessibility and reliability work continues.

Student-led, with formal teacher review planned. SpellTag is being prepared as a public-facing platform for the Hong Kong ICT Awards Student Innovation Award.

PRODUCT CHANGELOG

How SpellTag has changed since the first prototype

A concise release history of what the team built, what changed after Open SPACE and what is being upgraded for competition.

  1. Current buildCompetition upgrade

    Growing the demo into a fuller learning platform

    The competition build adds short spelling phrases, an installable PWA foundation and an explainable Learning Journey. Dictionary and admin workflows are still being strengthened before submission while the existing NFC and Wi-Fi hardware stays unchanged.

  2. After Open SPACEPriorities refined

    Showcase feedback shaped the next release

    Feedback highlighted the restrictive input limit and dictionary accuracy. The team also prioritised mobile access and clearer ways for children and adults to revisit each creation together.

  3. Original prototypeA+ final-year project

    The core Spell → See interaction worked end to end

    NFC alphabet cards, an ESP32-S3 and PN532 reader, Wi-Fi and web-based AI image generation formed the first complete SpellTag system.

STUDENT-BUILT, EDUCATOR-GUIDED

A student team growing SpellTag with educator guidance

SpellTag began as an A+ final-year project and is now being upgraded for more reliable, accessible and supervised learning use.

Student product team

Hardware, platform, AI pipeline and experience design.

Academic review

Technical evaluation and an A+ final project result.

Next pilot

A small educator-guided pilot is planned with adult consent, clear observation goals and transparent limitations.

Review safety and project status

READY TO SPELL SOMETHING?

Build a word and see what the AI imagines.

Try curated demo Create a new picture
The demo uses curated examples. New images come from the controlled creation service and should be explored with adult guidance.