Obscura 0.2.0 introduces :efficient, combining deterministic PII recognizers with spaCy NER for English names and locations.

Inference runs in a local Rust executable through an Erlang Port, with supervised workers and no Python required at runtime. It supports Linux servers and Apple Silicon macOS development.

On 5,000 heldout synthetic documents, it recovered more exact PII spans than :fast, but also produced more false positives. The article explains that tradeoff, CPU workload results, and why higher F1 doesn’t necessarily mean better privacy protection—with installation and Elixir integration examples.

Feedback and synthetic failure cases are welcome!

https://hfiguera.github.io/obscura/blog/beyond-regex-why-we-shipped-a-cpu-profile-for-pii-detection/