Maria

• 

  Today’s subscription screen

• 

  Today’s real iPhone screenshot analysis and Danger reply

• 

  Uncropped app icon

Inspiration

Our parents deserve help figuring out the internet, too. Especially the part where one wrong click can cost real money.

I'm Gabe, a solo founder in Rio de Janeiro. I built Maria to help our parents find their way online, including DeFi. When a message feels confusing, they should have somewhere to ask about it before acting.

What it does

Maria is an AI assistant for iPhone. Share a suspicious message or screenshot, ask what it means, and keep asking questions about the parts you do not understand.

For example, a text claims you won $2,000 and asks for $50 to claim it. Maria can identify warning signs, explain the risk in plain language, and suggest what to check next. Her answers can contain mistakes. She cannot move money or connect to a wallet.

How I built it

I used Expo, React Native and TypeScript for the iPhone app. Replit hosts the server. With the user's permission, OpenAI processes the conversation and selected images. Clerk handles Apple sign-in.

RevenueCat powers the Monthly subscription and Restore purchases. Server-side verification connects a valid purchase to the user's usage allowance.

Challenges

Testing account deletion exposed a missing return path. I added an explicit restart flow and recorded it working with native Apple sign-in on an iPhone 15. I also checked the live chat and screenshot flow on that device before resubmitting build 16 to Apple.

What I'm proud of

Maria can take someone from a confusing screenshot to an explanation they can question. In September 29 sandbox tests, a Monthly purchase granted one usage allowance. Restore on the same signed-in account preserved it without adding duplicate credit.

What I learned

A reply needs to help someone decide what to do next. I built the conversation so they can ask about anything that remains unclear. Testing the installed iPhone app also exposed issues that browser checks did not catch.

What's next

I want to train open-source models with LoRA and QLoRA using opt-in examples, with evaluations before each release. Contextual embeddings and retrieval-augmented generation (RAG) could help Maria find relevant, trusted guidance.

With funding, I would hire ML and trust and safety specialists, invest in GPU infrastructure, add languages, and bring Maria to the web. These are future plans. Today's app uses OpenAI through Replit.

Built With

• apple
• clerk
• expo.io
• openai
• react-native
• replit
• revenuecat
• typescript