PrivAiTe: Self-hosted proxy that redacts PII from LLM calls, incl. tool-calls (github.com)

PrivAiTe has launched a groundbreaking self-hosted proxy designed to enhance privacy when using large language models (LLMs) by reversibly redacting personally identifiable information (PII) during API calls. This innovative tool sits between a user’s application and the LLM provider, automatically replacing PII such as names, emails, and phone numbers with placeholders before any sensitive data leaves the user's environment. The system operates without telemetry, ensuring that no personal data is sent online, while also seamlessly handling PII within tool-call arguments and multimodal content—areas where many existing solutions falter. The significance of PrivAiTe lies in its dual detection engines: Presidio, which employs regex and basic named entity recognition (NER) for structured PII, and a more advanced Privacy Filter that leverages machine learning for context-aware detection. By integrating these two systems, PrivAiTe provides improved accuracy with a recall rate of 84.5% while maintaining minimal false positives. Users can opt for different presets based on their needs for speed or thoroughness, making it a flexible solution for developers concerned about data privacy, particularly in compliance with regulations like GDPR and HIPAA. This advancement marks a critical step toward greater control over PII in AI applications, emphasizing the importance of data sovereignty in the AI/ML community.