Security pipelines for LLM applications using ONNX models from Hugging Face Hub. Detect prompt injections, jailbreaks, and PII leaks. Models are lazily downloaded and cached locally. Fast local inference (~10-20ms after initial load).

Secryst (scrying + crystal) reveals the hidden reading of a script — diacritization, vocalization, grapheme-to-phoneme — via local, sha256-verified IMF v1 ONNX models. Implements the interscript-ml contract (models.yaml index, byte tokenizer, golden parity). Sibling crystals: pip install secryst, npm i secryst.

Ruby implementation of Google's SigLIP2 model for creating text and image embeddings. Uses ONNX models from HuggingFace onnx-community.