What problem does it solve? Processing large archives (email dumps, transcript libraries, document collections) with a single model forces a bad tradeoff: a cheap model misses nuance on important content, while an expensive model on everything costs 10-50x more on a corpus that is mostly noise. ## Core Features & Use Cases - Tiered model routing: A utility-tier model triages and classifies each item in about 2 seconds, a gate escalates high-value items to the deep tier, and the reasoning tier handles the default deep read. - Deterministic privacy wall: Named-entity and sensitive-pattern matching strips or diverts private content before any LLM call, with ambiguous items failing closed to human review. - Single-pass pipeline with checkpointing: Triage, gate, deep read, and immediate brain-page write happen in one pass with progress saved every 25 items for crash resilience. - Use Case: Processing a 16,000-item email archive where roughly 70% is noise — triage everything cheaply, deep-read only the significant threads, and cut cost roughly in half versus running the best model on everything. ## Quick Start Use two-tier extraction to process my email archive so only the significant threads get the deep-tier model.
