Lessons Learned Building A GenAI Powered App
summary
Case study of an AI-powered trivia app that generates quizzes with Gemini Ultra and images with ImageGen v2, running on Google Cloud Run (Flutter UI, Flask API, Firestore realtime, Vertex AI). Generation is fast but image calls take ~3–10s; hallucinations and malformed JSON happen regularly. Validation pipeline shows improvements: validator: Palm ~80% → Gemini ~91%. Engineering fixes: pin model versions, version prompts/parsers, batch and parallelize calls, cache outputs, and replace LLM scoring with fuzzy matching where cheaper and more reliable.
AI-written summary · about
chapters (11)
- 0:00Introduction and context
- 1:05App origins before GenAI
- 6:04GenAI arrives and possibilities
- 8:57Live demo of the quiz app
- 13:10Application architecture overview
- 18:31GenAI quiz generators evolution
- 22:26New challenges with GenAI
- 25:32Key lessons learned
- 34:47Coding defensively with LLMs
- 41:12Testing and validation strategies
- 45:25GenAI unblocked new use cases
related talks
01h 01mLessons Learned Building a GenAI Powered App
03mAI generated trivia games on Cloud
01h 11mBuild Your Own AI Sidekick with Flutter
30mGenAI on Android with Gemini Nano and Firebase
36mEmpowering Your Flutter Apps with AI
01h 03mBuild & deploy AI-powered apps
38mGenerative UIs: Using LLMs to Create Flutter Screens at Runtime
14mPrompt to production: Building an AI app with Flutter