How I put AI inside three live apps

I have three apps of my own in the stores: GiftPlannerAI, ToastAI and Famtale. All three are built with React Native and TypeScript on a Firebase backend, and all three use AI for the main job, not as an extra button. Each one uses AI in a different way, and each taught me something.

GiftPlannerAI: AI that keeps working

GiftPlannerAI reminds you of birthdays and suggests gifts, cake designs, messages and postcards. Gemini writes the ideas. If Gemini fails, Groq takes over automatically, so the user still gets an answer. There is also a chat assistant that can open the right screen in the app for you.

Privacy rules are written into the code. Friends' names never go to the image generator. The assistant asks for consent before it sends anything to an AI service. Sensitive guesses about health, religion or politics are filtered out. Daily and weekly limits keep AI costs predictable.

ToastAI: your words, in your voice

ToastAI writes wedding and birthday toasts and reads them aloud. Gemini writes the toast, again with Groq as an automatic fallback. ElevenLabs reads it in the user's cloned voice, and Whisper turns a spoken voice note into a first draft.

A toast is full of personal details: names, memories, inside jokes. Those go only to the AI on the server, where the API keys live, and never to an image service. In-app purchases are checked on the server too.

Famtale: one family, many languages

Famtale asks the whole family one question a day. A family often speaks more than one language, so every text and voice answer is transcribed and translated for each reader. This runs on Gemini through Vertex AI, where family data is not used to train models.

The rule "answer first, then read" is enforced by Firestore security rules on the server, so no one can read the family's answers before sharing their own. Every Sunday evening, Cloud Functions turn the week's answers into one chapter of a family memory book, translated for every member.

What I learned

• Every model needs a fallback. AI providers have outages and rate limits. A second provider means the user still gets an answer.
• Keys and prompts belong on the server. The app never holds an AI key, and personal details are sent only where they are needed.
• Write privacy as code, not as a promise. Name filters, consent checks and security rules work even when someone forgets the policy.
• Put limits on AI from day one. Usage limits keep costs predictable, so a free app can stay free.

If you are planning an app with AI at its core, tell me about it.