multi-modal AI to avoid eating takeaways
summary
Built an Android Kotlin Multiplatform app that sends an image and prompt and returns a structured grocery list plus a recipe, implemented with Kotlin Multiplatform Compose, Ktor, Kotlin serialization and Base64-encoded images; uses Gemini Flash via Vertex AI REST (Firebase routes an alternative), handles token regeneration and brittle JSON outputs, and notes Gemini Nano is on-device only for Pixel 8/Samsung flagships. Demo mostly works, common failures are access-token expiry and malformed JSON; cost measures ≈0.001 US cents per call.
AI-written summary · about
chapters (11)
- 0:00Multimodal AI for dinner ideas
- 1:08Multimodal app with Gemini LLMs
- 3:57Vertex AI for API access
- 6:47Testing LLM prompt with dish
- 7:51Live demo of recipe app
- 12:11Kotlin dev tools and frameworks
- 17:50Debugging prompts and API
- 19:39Vertex AI cost analysis
- 23:30Server-side LLMs dominance
- 25:42Next steps and improvements
- 26:25Q&A
related talks
28mDinner for two with Gemini: takeaway or restaurant?
35mBuilding Androidify: an AI-powered Android experience
30mGenAI on Android with Gemini Nano and Firebase
41mGen AI for Android Developers
21mAdd Generative AI capabilities to your Android app
01h 11mBuild Your Own AI Sidekick with Flutter
39mGetting Started with GenAI for Android Apps
14mPrompt to production: Building an AI app with Flutter