An AI Engineer Built a Communication App for His Nonverbal Autistic Son — Speech Went Up 5x, and It Accidentally Became a Small Business

When a Child Won't Speak, Every Pitfall a Parent Has to Clear First

A child not speaking is hard to notice at first, because no child speaks from birth. So a year or two just slips by. By the time you and your spouse realize your kid should be talking and isn't, the pediatrician often just says "it's fine."

Then you keep noticing something's off, the doctor starts frowning, out come the diagnostic questionnaires. You fill them out, then talk yourself into believing you're just being too pessimistic, that the evidence isn't really there. Until one day you watch a kid the same age as yours carrying on a full back-and-forth conversation right next to your child, and you can no longer deny it: something really is wrong.

Next comes an entire pit of "so now what." A bunch of quacks will surface, trying to sell you things that are completely unfounded.

Once you've knocked on every door worth knocking on, you eventually land on a speech therapist. This step genuinely helps — therapists know a lot of things you don't. But what they can do is still limited; a lot depends on the child's innate capacity. If the child is deaf, mute, or has a physical impairment, therapists have plenty of ways to help them communicate. But if the block is in cognition itself, all you can do is run drills over and over and see whatever progress comes.

Why Off-the-Shelf AAC Devices Didn't Work for His Son

Speech therapy will often equip kids who struggle to talk with a device called an AAC. Put simply, it's a tablet full of words and symbols — you combine symbols into a sentence and the tablet speaks it aloud. Essentially it's an electronic version of "point at the picture to say it," barely a step up from a file directory.

The author's son used an AAC for about a year and never showed interest — he'd play with it for a few minutes then wander off to other toys. There was one at school and one at the therapy office; they never bothered adding one at home, because he simply wouldn't engage with it. Every therapist and teacher agreed: he wasn't interested in the tool, and there was no forcing interest.

The problem was the symbol system itself. A red octagon means "stop" — his son had no idea, he didn't even understand what "stop" meant in that context. Arrows, those tiny stick figures — same story, he couldn't read them. The exact thing AAC asks of a kid is the exact thing he struggled with most.

Swap Abstract Icons for "His Own Life"

The author's day job is AI engineering (he describes himself as a Forward Deployed AI Engineer). He sat down, used vibe coding (intuition-driven programming: describing what you want in conversation with AI and letting it generate the code, while you just steer and refine) to put together a simple website with basic navigation in two hours, then generated a few hundred vocabulary images in ChatGPT.

He loaded the site onto a touchscreen laptop at home and handed it to his son. The boy was instantly captivated. To him, this was his own life laid out like a picture book, like his favorite cartoon — he recognized himself in the images without any effort. He pressed the photo of his grandpa over and over, then said the longest sentence he'd ever spoken in his life: "I really love you a lot." It's a phrase his grandpa says to him all the time.

They photographed every food he had, and taught him that "the picture of a hand reaching out" means "I want," then taught him how to flip to the food list. For four years, the author had never been able to figure out what this kid wanted to eat — suddenly it became effortless: "I want orange," "I want peanut butter crackers," "I want French toast sticks." Which toy upset him? Tap that picture, problem solved. His son loves it when the author presses his forehead hard against his — so he made a dedicated button that says, in his voice, "I need a head squeeze" when pressed. The whole product was built around this one child. These small-seeming but massive wins made the author genuinely mad at himself for not doing this sooner.

Two Things Even the Off-the-Shelf Devices Lack: Data and "Teaching"

Once his son was fully engaged, the author started getting angry at the entire industry. He found that most AAC devices on the market are missing two absolutely basic things.

So the author added both: metrics, and a teaching mode that actively instructs the child. After two weeks, his son's ability to look at a word and pick the matching picture out of a group improved by more than double. What he built is this constantly-turning loop below.

He even gave the app a schedule: every so often it plays a song, and his son walks over on his own to play a matching game or watch a slideshow. The app asks him every 45 minutes if he needs the bathroom. On the drive to appointments, the author taps a few things on his phone and his son's tablet switches into "preview mode," specifically teaching him what he'll encounter at the clinic. When he wants to ask his son something and the boy doesn't get it right away, he types it into the phone app, and the text shows up on the boy's screen rendered in the symbol set he's familiar with — like hieroglyphics — and he understands instantly, because the meaning appears exactly where his attention is already focused.

After Finding His Voice, an Unhappy Kid

After three weeks, his son became a genuinely unhappy child — really angry. Because now he had a "voice," but wasn't used to having one yet. Being understood came with responsibility attached.

Before, when he was hungry or upset, it was just a vague mood his parents had to guess at. Now his parents point at the tablet and tell him, over and over, "tap it, say it" — making him express himself. For the first time, he had to take responsibility for "saying it." And the frequency of his speech went up too — roughly five times what it used to be.

That bite of waffle didn't come easy. The author had tried everything to get his son to eat more variety, including strict rules — nothing ever worked, about as effective as trying to give a cat medicine. This time, they made a few short pieces inside the app: a song about "how great it is to try other foods," paired with photos of him actually trying new foods, narrated in Dad's voice. And then, his son ate the waffle.

From "Built for My Son" to "Something Other Kids Can Use"

Now the author is thinking about what else to add to the app to help other parents. Turning a tool custom-built for one child into a product any family can set up in ten minutes means solving a pile of technical problems along the way.

For instance, he first needs to build a general-purpose classification system for early childhood vocabulary — a kind of "scaffolding." That way, when parents casually snap photos at home, the system can automatically sort them into fixed slots, which is the foundation everything downstream of automation runs on.

He wants the parent experience to be as simple as possible: tell the system your kid's favorite color, pick an animation style your kid likes, maybe read a 30-second script to record a voice, upload a few photos, and the communication board builds itself. He wants to absorb all the complexity himself, so that a mom juggling four kids and barely keeping her head above water can still get the whole thing set up in ten minutes.

There's an even harder problem: how do you quantify whether a given child has actually "mastered" something? It varies from kid to kid. The author's son has weak joint attentionThe ability to share focus on the same thing with someone else — for example, if you point at an object, the child follows your gaze to it. It's a foundational skill for language and social development, and many autistic children are weaker in this area.: if the author manages to hold his attention during a given test, he can get nearly everything right; if not, he'll just tap randomly until the test ends. The exact same set of questions can produce wildly different results.

His day job happens to involve something in a similar direction (he's in the final sprint before launch, which is also why he's been so busy lately). Once that's built, he plans to let parents talk to the app in plain language and have it adjust the interface in real time.

How Much It Costs, and Who Already Wants It

The kind of dedicated AAC device mentioned earlier can be expensive. Buying the hardware directly from a vendor can run over seven thousand dollars; if you already own an iPad, a subscription is reasonably priced.

By the author's math, his app can run at $9.99 a month and still cover the cost of running all the AI calls needed to make it work as well for other kids as it does for his own son; with voice cloning it's $19.99 a month, since keeping that feature always-on costs more.

His son's speech therapy clinic wants to use it with other kids; so does the school. The author has never pitched it once — which is what he means by "accidentally starting a small company." There's plenty of red tape to sort through too — for instance, he can't call it a "medical device," and he can't call it an "AAC," so in the marketplace he can only describe plainly what it does, and sell it that way.

His goal is that once parents no longer need to keep generating new images, he'll make it free, or at least keep a bare-bones free tier — because once a child has a few thousand images generated, there's not much left to generate. That's the lowest price he can responsibly offer, since once it's live, he's on the hook to keep the service running for every single child using it. He says in a few more weeks, other families will be able to try it.