I’m a developer in transit. I’m constantly switching between coding courses and boot camps when I hit roadblocks. In my defense, linear learning is hard when you’re neurodivergent.
But artificial intelligence makes it easier to iterate until I nearly reach perfection. It’s why I was confident going into vibe coding, and how I wound up building a breast self-exam app for mobile devices.
Building mobile apps with AI
Trivia challenge
From code generation to on-device models — find out how much you really know
about AI-powered mobile development.
AI ToolsFrameworksOn-DevicePlatformsConcepts
Which AI-powered coding assistant, developed by GitHub and OpenAI, became one of the
most widely adopted tools for generating mobile app code?
Correct! GitHub Copilot, launched in 2021 and powered by OpenAI’s Codex
model, quickly became the go-to AI pair programmer for mobile developers. It integrates directly into
editors like VS Code and can generate entire functions or UI components from a simple comment.
Not quite — the answer is GitHub Copilot. Launched in 2021 through a
partnership between GitHub and OpenAI, Copilot uses the Codex model to suggest code in real time. While
tools like Amazon CodeWhisperer and Tabnine are strong competitors, Copilot was the first to achieve
widespread mainstream adoption.
What is the name of Google’s on-device AI runtime that allows machine learning
models to run directly on Android devices without a network connection?
Correct! TensorFlow Lite is Google’s lightweight version of TensorFlow
specifically designed for mobile and embedded devices. It lets Android developers run inference
on-device, which means faster responses and no data ever leaving the user’s phone.
Not quite — the answer is TensorFlow Lite. Google built TensorFlow Lite
to strip down its full TensorFlow framework for mobile hardware constraints. Core ML is Apple’s
equivalent for iOS, while ONNX Runtime is a cross-platform option, but TensorFlow Lite has been the
dominant Android-native solution.
Which Apple framework allows iOS developers to integrate trained machine learning
models into their apps and runs exclusively on Apple devices?
Correct! Core ML is Apple’s dedicated machine learning framework for
iOS, macOS, watchOS, and tvOS. Introduced in 2017, it allows developers to drop in trained models and
run inference on-device using Apple’s Neural Engine for maximum efficiency and privacy.
Not quite — the answer is Core ML. Apple introduced Core ML in 2017 to
give developers a unified, hardware-accelerated way to run ML models locally. ARKit is Apple’s augmented
reality framework, and while it uses AI internally, it isn’t a general-purpose ML integration tool for
developers.
In the context of mobile AI development, what does the term ‘model quantization’
primarily refer to?
Correct! Model quantization reduces the memory and computational
footprint of a neural network by representing its weights with lower-precision numbers, such as 8-bit
integers instead of 32-bit floats. This makes models far more practical to run on resource-constrained
mobile hardware.
Not quite — the answer is reducing model size by lowering the precision
of its weights. Quantization is a key optimization technique for mobile AI because smartphones have
limited RAM and processing power. A quantized model can run several times faster and consume a fraction
of the storage compared to its full-precision counterpart.
Which cross-platform mobile framework introduced ‘AI-assisted UI generation’
features through integrations with tools like Firebase Genkit and Vertex AI?
Correct! Flutter, Google’s cross-platform UI toolkit, has been expanding
its AI capabilities through integrations with Firebase Genkit and Google’s Vertex AI platform. Google
has positioned Flutter as a key framework for building generative AI-powered apps across Android, iOS,
and the web from a single codebase.
Not quite — the answer is Flutter. Google has been actively pushing
Flutter as the framework of choice for AI-integrated apps, particularly through its Firebase Genkit and
Vertex AI tooling. While React Native also has AI integrations available, Flutter has received the most
direct investment from Google in this specific area.
What is Prompt Engineering in the context of building AI-powered mobile apps?
Correct! Prompt engineering is the practice of carefully designing the
text inputs sent to a language model so that its outputs are accurate, consistent, and useful. For
mobile app developers using APIs like OpenAI or Gemini, strong prompt engineering is often the
difference between a polished AI feature and an unpredictable one.
Not quite — the answer is crafting structured input text to reliably
control the output of a language model. Prompt engineering has become a critical skill for mobile
developers integrating large language model APIs into their apps. Rather than modifying the model
itself, developers shape its behavior entirely through the instructions and context they provide in the
prompt.
Which chip component, found in modern smartphones like the Google Pixel and Apple
iPhone, is specifically designed to accelerate AI and machine learning workloads?
Correct! The Neural Processing Unit, or NPU, is a dedicated hardware
block engineered to handle the matrix multiplication operations that power neural networks. Apple calls
theirs the Neural Engine, Google calls it the Tensor Core, but both serve the same purpose: running AI
inference quickly and efficiently without draining the battery.
Not quite — the answer is NPU, or Neural Processing Unit. While GPUs can
also accelerate ML workloads, NPUs are purpose-built for the specific mathematical operations used in
neural networks, making them far more power-efficient for sustained AI tasks on a phone. The ISP handles
image signal processing, and the DSP handles general digital signal tasks.
When a mobile app sends user data to a remote server to process an AI request and
return a result, this approach is commonly called what?
Correct! Cloud inference means the heavy lifting of running an AI model
happens on remote servers, not on the device itself. This allows apps to use massive, powerful models
that would never fit on a smartphone, but it comes with tradeoffs including latency, internet
dependency, and potential privacy concerns around user data leaving the device.
Not quite — the answer is cloud inference. When an app sends a request
to an external API like OpenAI’s GPT-4 or Google’s Gemini, that is cloud inference in action. On-device
inference is the opposite approach, keeping all processing local. Edge computing sits somewhere in
between, using nearby servers rather than distant data centers.
Your Score
/ 8
Thanks for playing!
There are no malignant bugs in my system
Only clean, healthy code allowed here
My sister is a breast cancer survivor. Her surreal healing journey pulled our family into her orbit, and forced us to recalibrate our lifestyles.
Before, the illness was something that happened to other people. Experiencing it up close changed how I think about health.
Since then, I’ve encouraged my sister to document everything, especially her pursuit of a natural cure. It mattered enough to both of us that we kept returning to the idea.
We wanted to offer hope to others going through the same thing. Yet, everyday life kept pushing it down the list.
Then Simple Design’s My Calendar app gave me a reason to revisit the original e-book concept and improve it.
I’ve used the Android app for years because it’s simple, syncs across devices, lets me share cycles with my partner or track a friend’s, and makes it easy to send my history to my gynecologist.
Recently, it added a breast self-exam feature. I grinned when I saw it because most period apps focus on ovulation and conception.
The feature is bare-bones, but a recent reorganization gives each symptom its own entry. I’m a sucker for tiny details in apps I often use, so this small change makes a huge difference.
Braska is concerned about your everyday wellness
Because everyone deserves dedicated health tools
Women’s health remains largely underfunded and underresearched, and breasts as a topic continue to receive shaming.
A 2025 Novartis and The Harris Poll survey of over 3,000 US women revealed society influences more than half of women’s views about breasts.
One in three women is uncomfortable discussing their health outside a doctor’s office, and only one in three performs monthly self-exams.
I may not close the research gap, but I can create a solution that allows people with chest tissue to check in on themselves daily and freely.
The idea, at its simplest, is a companion similar to a period app. Its name is a play on “bra” and “asker,” yet tentative while I ponder a proper way to represent broader user groups.
You’ll enter what you noticed after each exam, and it compares your answer with past logs. Separately, there’s a gallery of common variations for you to compare symptoms against.
The Future Me journal is where you reflect on your history when anxiety makes new sensations frightening. You may document tenderness you’ve experienced before and what happened afterward.
It reminds you of harmless patterns and offers some reassurance. You don’t want health scares affecting your quality of life before you seek professional consultation.
I built my app with React Native and Expo
It features interactive features and 3D anatomy
I needed the best AI for my sister’s app, so I let Claude, ChatGPT, and Gemini weigh in. ChatGPT and Gemini recommended Claude, while Claude gave a more nuanced answer about the strengths of each model.
So, I used ChatGPT and Gemini as my design directors. I turned their suggestions into a project brief, which I then gave to Claude.
Claude generated two versions of Braska. The first was a clickable React web prototype that I previewed in Chrome. It proved that the concept was largely feasible.
I moved to the three horsemen of my unglamorous developer journey: React Native, Expo, and TypeScript.
React Native turns the code into an iOS or Android app, Expo handles much of the setup, and TypeScript catches coding errors as I build.
I focused on the guided exam and adding a haptic pressure guide. It pairs light, medium, and firm modes with an animated breast ring and matching vibrations.
Optional voice narration exists through Expo Speech and lets you complete the exam hands-free. The AR mirror uses the front camera as a live reflection with guides showing where to examine.
I deliberately skipped pose detection and computer vision to ensure the app never captures or stores images.
The 3D anatomy model was my favorite part to discuss with Claude because the AI generated the model in Three.js from basic shapes.
So, visualize spheres for lobules, cylinders for ducts, and smaller shapes for the nipple, areola, and lymph nodes. You can rotate it and tap structures to learn what they do.
You have to understand that anatomical accuracy is a much higher bar than making something look like a body part.
Although I wouldn’t treat this as a medical-grade model, I love being able to create it from primitives rather than sourcing a model from somewhere else.
It’s the best way for me to sidestep asset licensing concerns and match every structure to Braska’s color palette.
Your privacy should come first on any app
Encryption-level protection is my standard
Braska encrypts user data with XChaCha20-Poly1305. It generates a key on first launch and stores it in the phone’s secure storage through iOS Keychain or Android Keystore.
Any corrupt data becomes unreadable rather than falling back to plain text. Claude’s code was impressively clean from what I could verify, but I took time to review the UI copy.
The last major hurdle was updating my MacBook to macOS Tahoe 26.6.2 so that I could install the Command Line Tools.
A previous Apple update wrecked my old Intel Mac, so I checked forum feedback before upgrading. This time, it went smoothly.
I then installed Homebrew, using it to install Node and later Java 17 when Android Studio’s newer Java version proved incompatible with my React Native project.
Then I created a virtual Google Pixel 10 and ran my first build after installing Android Studio. It took more than 10 minutes with nothing cached, but Braska ran successfully.
There are rough edges I need to fix. For one, the AI check-in currently only follows pre-written scripted responses and can’t handle user input yet.
Regardless, I’m thrilled to have gotten this far. With more testing and debugging, the next step is putting the code under version control and pushing it to GitHub.
Good vibes make better code
Coding is less intimidating as AI handles more mechanical work. It’s capable enough that you can even create widgets with tools like Gemini.
Yet, it doesn’t undermine developer efforts behind app building. You always need to have a clear sense of what you’re trying to achieve.
Otherwise, you’ll make a fully functional app with no compelling reason to exist. So, don’t lose sight of your software’s purpose between prompts.
