Last week, I set out to build an AI art generator for Axel. Simple goal. But then I thought—what if he could also use it to learn to read? And down the rabbit hole we went.
Axel struggles with prefrontal synthesis (PFS)—the ability to mentally visualize objects in different positions. To support his language processing, I structured the app’s sentences to be visually concrete and spatially clear:
[Subject] + [Verb] + [Object] + [Preposition] + [Location]
“Axel holds toy above table.”
This follows MITA therapy research, which shows that structured spatial language strengthens cognitive development. Every word is predictable, processable, and drawable—so Axel can see what he’s saying.
But then: As you can see from the demo, the LLM failed me.
I expected: “Axel holds…” → [“toy”, “cup”, “ball”]
I got: “Axel holds…” → [waiting.. waiting.. waiting for a broken phone] 😭
So, like any good neurodivergent, I scrapped the LLM and went full DIY. Now, instead of using an external API, I’m working on:
✅ Training a custom FastText model on a finite, controlled vocabulary
✅ Ensuring only context-appropriate words appear as suggestions
✅ Making word selection lightning-fast (because LLMs can be slow as hell)
✅ Avoiding “destiny” and other poetic nonsense in Axel’s AI prompts
This means more control, better accuracy, and instant response times—because the last thing I need is Axel rage-quitting his own learning tool.
Scope Creep? What Scope Creep?
Was this supposed to just be an AI art generator?
Yes.
Am I now training a custom word prediction model, designing a structured language-learning tool, and integrating it with assistive tech?
Also yes.
Do I regret it?
Absolutely not.
Because at the end of the day, Axel will have something that actually works for him. A system that helps him learn to read, build sentences, and create art—all on his terms.
Meanwhile I have generated up some curated words so we can get this party started.
#AssistiveTech #AI #Autism #Accessibility #LanguageDevelopment
