Your ML portfolio should demonstrate end-to-end thinking, not just model accuracy. Hiring managers look for: real data (not Iris/Titanic), clear business problem, production considerations, and documented tradeoffs. Top 5 portfolio projects: 1) RAG-powered chatbot for a specific domain (medical Q&A, legal research) - demonstrates embeddings, retrieval, prompt engineering, and evaluation. 2) Recommendation system with implicit feedback - shows collaborative filtering, handling cold start, A/B testing mindset. 3) Real-time ML inference API - demonstrates MLOps, model serving, monitoring, Docker/Kubernetes. 4) Document intelligence pipeline - OCR, NER, classification, showing end-to-end data pipeline thinking. 5) Time series forecasting with business metrics - demand forecasting or financial prediction with proper backtesting and business framing. How to present: lead with business impact (not accuracy%), explain the data and why it is challenging, discuss what failed before the current approach, address production considerations. GitHub README structure: problem statement, demo link, architecture diagram, key decisions and tradeoffs, performance metrics, how to run locally.