My idea was to create a device that could use AI in a way that’s both engaging and educational for kids. I wanted to replicate the “chat interface” experience I get with tools like ChatGPT, but in a simple format tailored for younger users. Something like a phone device that they could have “chat” conversations with while practicing reading and writing.

By combining portable hardware capabilities with the power of OpenAI’s Assistants API, I saw an opportunity to create a unique learning tool. This device wouldn’t just answer questions; it would engage my sons in thoughtful conversations, encourage follow-up questions, and adapt to their individual communication styles and interests.

The goal was to create more than just another gadget. I wanted a tool that would spark curiosity, improve their reading and writing skills, and teach them how to ask good questions. I wanted to train them to use AI as a tool, because I see that as an important skill going forward.

After considering various options, I settled on the LilyGo T-Deck (docs) as the hardware platform for this project. My previous experience with the LilyGo T-Screen had been positive, and the T-Deck offered several advantages that made it ideal for a portable AI device. It actually comes WITHOUT a case, but that gave me a lot of flexibility when it came to the final design.

The integrated keyboard is a key feature for this project. It allows the kids to type their questions and responses naturally, improving their typing skills while interacting with the AI. The touchscreen adds an extra layer of interactivity, making the device feel more engaging and modern.

The microSD card slot is another critical component. It allows for local storage of conversation threads, enabling the device to maintain context across multiple sessions without relying on constant internet connectivity.

At a retail price of about $50 USD, the T-Deck offers an excellent balance of features and affordability. This makes it a viable option not just for personal projects, but potentially for broader educational applications.

While the T-Deck isn’t specifically designed as a children’s device, its robust construction and simple design make it well-suited for this application. The addition of a custom 3D-printed case (discussed later) further adapted it for younger users.

“You are a knowledgeable and friendly AI assistant for middle school-aged children. Your responses should be informative but age-appropriate, encouraging curiosity and critical thinking. Avoid complex jargon, and always be patient and supportive. If asked about sensitive topics, provide factual, neutral information suitable for this age group.”

This is just a snippet of my custom prompt but it serves as a foundation, guiding the AI to interact in a way that’s engaging and educational for my young users.

It’s a delicate balance – I wanted to encourage learning and exploration, but also ensure the content remains appropriate and understandable for middle schoolers. If it uses words beyond their comprehension level, they wont use it. It has to be fun and speak directly to them.

One of the most exciting aspects of using the Assistants API is that by using threads and the Assistant prompt I limited the payload and memory requirements. By maintaining conversation threads on OpenAI’s servers, I significantly reduce the amount of data we need to send with each interaction. This not only keeps costs down but also improves response times, making for a snappier, more enjoyable user experience. Having to pass just the next question back each time also allows me to get the interaction working on this low memory device.

The software I wrote has really just 2 main screens written in LVGL. The main screen shows the “New Question” button and the list of previous threads. New questions create a thread_id on the server and a file on the SD card. Conversation threads can be continued by clicking into them or edited by clicking the pencil (edit) button. Essentially this screen is simply a list of buttons with the top button always being the “New Question” buton. When a new question is added it creates a new button on this list.

The second screen is more of the conversation screen. It is used for both the new question and continue conversation ui. It has a textbox to send a new message to the thread as well as the thread messages laid out in a chat configuration so that the boys can tell the difference between their messages and the OpenAi Assistant response messages.

You type in your initial question and hit send. The device shows a loading dialog while it sends the request to OpenAI’s Assistants API. The first request starts the thread and the thread endpoint is polled until it has answered the question. The full request and response takes about 6 seconds. As you add additional questions the device knows what thread you are on so it will know what thread to poll for a response. list.

Because of how the Assistants API works, on the submission of a new question responds with the new thread_id but does not contain the question response. I actually had to write a polling method that would retry to query the thread until the thread’s “run” had a status of “complete”. Only after the run was complete could I fetch the actual response to the question. Once the response it retrieved it is loaded onto the screen and it is ready for further questions.

When it came to housing this AI-powered learning device, off-the-shelf cases for the LilyGo T-Deck just wouldn’t cut it. My young users needed something special, so I dove into creating a custom 3D-printed case that would be both functional and kid-friendly.

The design process, carried out in SketchUp, was an exercise in balancing competing priorities. We needed a case compact enough for children’s hands, yet durable enough to survive inevitable drops.

I designed a case, even though there are many different existing designs for the LilyGo T-Deck. I really enjoy spending time in sketchup and wanted to make something that was as small as I can so that the smaller hands wouldn’t feel that it was a burden.

I have a new BambuLabs P1S and decided to print in PLA. It’s cost-effective, easy to print with great detail, and tough enough for our needs. But the real fun came with color selection. I let my sons choose, resulting in one matte army green unit and one in sleek matte black. Nothing like a personal touch to get kids excited about technology!

The case’s technical specifications reveal where the real engineering challenges lay. I squeezed in a 2000mAh battery, striking a balance between longevity and compactness. The power switch got a kid-friendly upgrade, now featuring a satisfying click that serves as a tactile reminder to turn the device off. A custom reset button extension was also added for those “have you tried turning it off and on again?” moments.

One of my space-saving tricks was relocating the speaker to the side, allowing me to reduce the overall depth and snugly fit the battery. The entire assembly is held together with four inset hex 2.5M 12mm screws in black. The screws are completely recessed to prevent scratching surfaces, while providing rock-solid structural integrity.

The case features rounded edges for safety, a textured grip surface to prevent slips, and a slightly raised bezel to protect the screen. Clear simple controls ensures intuitive use, crucial for my young tech enthusiasts.

The result? We’ve transformed the developer-oriented T-Deck into a robust, kid-friendly learning tool. It’s not just a device; it’s a personalized tech companion that my children are genuinely proud to use and show off to their friends. Who says educational tools can’t be cool?

Did I hit my goals? I think so. Both boys regularly ask for their PocketGPT’s. The youngest for answering basic questions (mostly about foods) and the other tends to be asking questions about how to get past levels in his video games. They both kind of see it as their “device” since they dont yet have phones.

In the end, this OpenAI integration is what breathes life into our little handheld device. It’s not just about having a powerful AI model; it’s about how we’ve tailored and implemented it to create a truly unique learning tool. Who knows? We might be nurturing the next generation of AI enthusiasts right here!