Replacing NLP Pipelines with LLM Agents: MIS v2 Architecture

For years, the standard NLP pipeline looked like this: spaCy for NER, Hugging Face for classification, a custom regex layer for extraction, and a rules engine for routing. It worked, but it was brittle. Every new document type required new rules. Every edge case demanded a new training set.

The MIS v2 architecture replaces that entire stack with a Batch-and-Specialize pattern powered by LLM agents. Instead of a dozen microservices each doing one NLP task, a single Gemini agent with a 1M context window processes entire documents in one pass — classifying, extracting, and routing in a single inference call.

The result: 40% faster throughput, half the code, and dramatically lower maintenance burden.