AI startup technical decisions made early have long-lasting impact. MVP architecture: do not build your own LLM infrastructure - use OpenAI or Anthropic APIs. Focus on building the unique application layer. Standard stack for LLM product MVP: Next.js frontend + FastAPI/Node backend + Supabase (postgres + auth + vector) + OpenAI API + Vercel/Railway deployment. This is the fastest path to working product. Common MVP mistakes: 1) Premature optimization - do not build complex ML pipeline before validating product-market fit. 2) Over-engineering - avoid Kubernetes, microservices, custom vector databases at MVP stage. 3) Ignoring prompt engineering - 80% of quality improvements come from better prompts, not better models. Scaling decisions (at $10K+ MRR): move to dedicated LLM provider accounts, add Redis caching, add proper monitoring. At Series A: consider self-hosted models for cost optimization, build custom fine-tuned models if you have sufficient data, invest in proper MLOps. Model selection philosophy: default to Claude Sonnet or GPT-4o for quality. Add claude-haiku/gpt-4o-mini routing for cost optimization when you can profile which queries need quality vs speed. LLM provider diversity: use at least 2 providers with fallback logic. Single provider dependency is existential risk. IP protection: fine-tuned models and custom datasets are defensible IP; raw LLM API usage is not.