Dinner for two with Gemini: takeaway or restaurant?
summary
Two small apps actually get built and demoed: one converts a takeaway photo into a recipe and shopping list using Gemini multimodal and cloud JSON schema output, the other searches saved Google Maps places (Data Portability API + Places API stored in Room) and ranks matches. Implementation uses Gemini cloud + Google AI SDK and a Gemma 3 on‑device 1B model (≈350 MB) via MediaPipe, plus Ktor REST token hacks. Results work but are hacky: structured output breaks, retries and hallucinations occur (recipes claiming kilograms of cucumber), and testing costs range from €355 total to sample calls of ~$0.06. JSON schema structured output
AI-written summary · about
chapters (25)
- 0:00Welcome
- 1:03Saved lists problem
- 2:07Google AI SDK Android
- 3:33Fetching place details
- 4:18Designing prompts
- 5:02Prompt and data connection
- 6:07Lunch app demo
- 7:58Cloud LLM demo
- 9:03Image recipe generation
- 10:47On-device LLM Gemma 3
- 11:31Local LLM MediaPipe
- 13:17Recipe prompt engineering
- 14:43Cloud vs on-device
- 15:49Production vs experimental APIs
- 16:53Gemini Nano limits
- 17:35Downloading Gemma 3
- 18:38Privacy benefits
- 19:20Handling hallucinations
- 20:26Fine-tuning with LoRA
- 21:30Structured output solutions
- 22:53Structured generation deep dive
- 24:18Publishing to Play Store
- 25:02User testing next steps
- 25:45Cost and monetization
- 26:51Q&A
related talks
27mmulti-modal AI to avoid eating takeaways
56mBuilding Multimodal Apps with Gemini 2 0 and Gemini SDK
01h 01mBuilding Multimodal Apps with Gemini 2.0 and Gemini SDK
30mBuilding AI-Powered Apps with NodeJS and Gemini | Armen Vardanyan #JSConfAM25
19mGemini 2.0: Beyond the Basics
29mGenAI + Google Maps
08mHow businesses can leverage Gemini AI
51mGemini, Google's Large Language Model