How AI Got Into My Apps — From API Calls to Product Thinking
A PM-turned-indie-dev's honest take on using AI in products: what worked, what didn't, and what actually matters
What I Actually Used AI For
My apps use AI in two distinct ways:
- Puppie / Chill Kitty: LLM-driven pet conversations + image generation API for custom pet portraits
- Building everything: Nearly all my code was written with AI assistance — I focused on deciding what to build, not how to type it
As a PM with no formal CS background, AI tools this past year genuinely made “from idea to App Store” possible for me.
Conversation API: Simpler Than Expected, Pricier Than Expected
Technically, plugging in an LLM is straightforward — maybe 50 lines of code. The real challenges are elsewhere:
- Prompt design: Getting the AI to reliably play a specific pet personality, stay in character, and not say anything weird
- Cost control: Every conversation burns tokens. More users = linearly higher costs
Puppie’s current LLM spend: ~$3.30/month. Manageable. But if DAU grew 10x, the math gets uncomfortable fast.
Image Generation: Half Delight, Half Headache
“Upload your real pet photo, get a cartoon version” — this was the most loved feature. Also the most painful to ship.
The problems:
- Output quality is inconsistent: sometimes magical, sometimes completely off
- Cost per image is much higher than text generation
- Users have zero tolerance for “it doesn’t look like my dog”
Current solution: generate multiple options, let the user pick. Imperfect. Good enough.
A PM’s Take on AI Features
AI isn’t magic. It’s a very capable intern with unpredictable output.
Using AI in a product well comes down to one question: What can AI do reliably, what does the user expect, and how do you close that gap?
Slapping AI into a product for the sake of it helps no one. But the right AI feature, in the right place, creates an experience that’s genuinely hard to build any other way.