Sentiment analysis has evolved from lexicon-based to neural approaches. Production options: 1) Fine-tuned DistilBERT: distilbert-base-uncased-finetuned-sst-2-english gives 91% accuracy on SST-2, inference <5ms on CPU. Best for high-volume, cost-sensitive use cases. 2) Custom fine-tune for domain: collect 1000+ labeled examples from your domain (product reviews, support tickets), fine-tune DistilBERT with Hugging Face Trainer, typically achieves 3-5% improvement over general model. 3) LLM-based classification (GPT-4o/Claude): highest accuracy, handles nuance and sarcasm better, but 100-1000x more expensive and slower. Use for: complex multi-aspect sentiment, ambiguous cases, when accuracy justifies cost. 4) Multi-aspect sentiment: "The product is great but shipping was terrible" - extract aspect-level sentiment (product: positive, shipping: negative). Use structured output with LLM for aspect extraction + sentiment. Real-time pipeline: streaming ingestion (Kafka) -> Faust/Flink for stream processing -> sentiment model inference -> aggregate to dashboard. Production tips: batch inference for throughput (HuggingFace pipeline batch_size=64), ONNX export for faster inference, quantize to INT8 with <1% accuracy loss.