Obscura is an Elixir library for detecting and anonymizing PII at application boundaries before data reaches logs, outbound APIs, or LLM prompts. It offers three detection profiles: a dependency-light :fast profile using regex/parsers for structured identifiers (emails, phones, SSNs, IBANs), and :balanced/:accurate profiles backed by local NER models for prose entities like names and locations. Key features include nested map/list redaction, configurable anonymization operators (replace, mask, hash, pseudonymize), Phoenix Plug integration, Logger helpers, and an LLM pseudonymization workflow using reversible session vaults. The post also covers benchmark F1 scores across datasets, deployment considerations for multi-GB model assets, and explicit security limitations — Obscura reduces PII exposure risk but is not a compliance guarantee.
