A developer shares technical design decisions made while prototyping an open-source educational mobile app using Claude Code. Key decisions covered include: choosing Claude Sonnet over Haiku or Opus for balancing cost, quality, and latency; switching from Supabase to Firebase after discovering Supabase was blocked in India; using single-input AI activity generation with strict JSON schemas instead of conversational chat; implementing two-layer machine translation (static i18n + Google Translate API for dynamic content); and optimizing for low-bandwidth users via image compression and text-only AI calls. The author emphasizes that AI-assisted development still requires critical human judgment at every step, especially for infrastructure and tradeoff decisions.
Table of contents
Table of ContentsPrerequisitesWhich Model to ChooseChoosing For Geography and CostChoosing the Programming Framework and Backend ArchitectureMachine Translation and Multilingualism“Create with AI” with Humans in the LoopOptimizing for Low BandwidthProducing a Demo VideoSummary2.4K Impressions1 Comment