How Cerberus ID turns a fraud score into approve, check and reject zones, and how calibrating against 10,806 real registrations raised auto-approval precision to 99.5% with 78% fewer false approvals.
SPAM Escudo filters contact-form spam in five stages, with the LLM last. Every LLM verdict teaches the cheap stages to catch more, and a weekly AI-assisted review of what still reaches the LLM took pre-LLM decisions from 55% to 78%, with the full stack above 90%.
How Transit Ledger processes invoices and receipts entirely on a local CPU server: a tiny classifier gates a 4B extraction model, code checks the totals, and real-world data exposed a shortcut the first model had learned.
How to take AI from experiment to production while keeping cost, accuracy and risk under control: a cascade that uses the right technology for each step, with real results from spam filtering, invoice processing and fraud detection.