byerlikaya/Septum
Privacy‑first AI middleware that anonymizes PII locally and only sends masked text to cloud LLMs when configured.
What's novel
Privacy‑first AI middleware that anonymizes PII locally and only sends masked text to cloud LLMs when configured.
Code Analysis
23 files read · 6 roundsA privacy-first AI middleware that locally detects and masks PII (names, IDs, emails, locations) in documents and chat messages before they reach cloud LLMs, then restores original values from LLM responses using session-scoped anonymization maps with coreference resolution.
Strengths
Genuinely deep multi-layer PII detection pipeline (Presidio + multilingual NER ensembles + optional semantic validation) with real algorithmic checksums (TCKN, IBAN ISO 7064), domain-aware span absorption, and thread-safe session management — all in a zero-network-dependency core package with clean Protocol-based dependency injection. The air-gapped zone architecture with queue-based bridge to internet-facing gateway is a well-thought-out security boundary.
Weaknesses
The detector.py file at 75KB is a monolith that would benefit from further decomposition; test coverage is solid for the core engine but the 37 recognizer files and 7 national ID validators have limited dedicated test files visible. The coreference resolution is a simple token-subset heuristic that may miss or over-match in complex multilingual contexts.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
93
Contributors
1
Files
183
Active weeks
2
Repository
Language
Python
Stars
2
Forks
0
License
MIT