Hello World

Inspiration

Most early-learning apps assume a child can read, tap through menus, or answer a verbal prompt. That leaves out a lot of autistic and ADHD kids aged 3 to 7, many of whom are pre-verbal or minimally verbal. And the games that do exist are often built around points, streaks, and leaderboards, which can overstimulate exactly the kids who need calm the most. On top of that, caregivers of these kids often carry a vague worry with no low-stakes way to turn it into an actual question for a health worker, short of a formal, stressful diagnostic process. We wanted to build something calm, private, and personal, using a child's own real room and real objects instead of stock photos or generic characters.

What it does

My World: a caregiver takes one photo of the child's room. Claude's vision model identifies objects that are safe and reachable for a young child (filtering out anything sharp, fragile, mounted, or otherwise unsafe), and the app turns those real objects into six different hands-on activities: find-it, matching, tracing, story sequencing, block stacking, and sorting by rule. Each one is scaffolded by a fading support-tier ladder, so a caregiver can start with full physical guidance and back off as the child is ready.

Worry to Question: a caregiver writes their concern about their child's development in their own words. The app rewrites it into one clear, neutral paragraph plus three fixed questions to bring to a health worker. It never screens, scores, or diagnoses; it just turns a worry into a conversation.

Both branches are built specifically for autistic and ADHD children aged 3 to 7.

How we built it

React, TypeScript, and Vite, shipped as an installable, offline-capable PWA so a session keeps running once it's generated. Claude handles both ends of the intelligence: Sonnet for the one call that reads a room photo, Haiku for text generation, both behind a single serverless proxy so the API key never reaches the client. Faces are blurred entirely on-device with TensorFlow.js and BlazeFace before a photo ever leaves the phone. There are no accounts and nothing is stored off-device. The whole visual language was designed from scratch to stay calm: no scores, streaks, or diagnostic language anywhere, checked by our own automated tests that scan the live UI for banned words. This program is built using the assistance of LLM.

Challenges we ran into

Getting the vision prompt right took real iteration: a photo of a real child's bedroom needed a strict five-part filter (portable, reachable, big enough to spot, safe to handle, not a person) to avoid surfacing furniture or anything sharp. Designing for an autistic and ADHD audience meant deliberately removing almost every default gamification instinct and replacing it with a quieter, matter-of-fact tone throughout. And partway through the build, two of us ended up independently redesigning the same screens, which meant carefully reconciling both without losing either person's work.

Accomplishments that we're proud of

A genuinely private pipeline, where face-blurring happens on-device before anything is sent anywhere, and nothing about the child is stored off it. Six distinct, hands-on games generated from one real photo of a real room, not stock content. A support-tier system that fades from full physical guidance to independent play on its own.

What's next

Expanding the object-recognition vocabulary, finishing the remaining polish on Worry to Question, and building out caregiver session history in more depth.