Whitepaper · v0.8
This technical note outlines Tamga's inline-path scanner architecture and its Shadow ML feedback loop. All benchmark data is public in the docs/benchmarks directory on GitHub and can be reproduced locally with make redteam-report.
A hard 5 ms ceiling on the inline path. Python NLP pipelines are banned from the hot path; a Go DFA plus SIMD scans at over 6 GB/s.
An Aho-Corasick DFA (coregx/ahocorasick profile), a SIMD prefilter, Luhn plus BIN/IIN validation and MERNIS national-ID checksums.
Unicode normalisation (NFC/NFD), Turkish diacritic folding, word-to-number conversion (EN + TR) and tokenless contextual proximity.
A UDS + FastAPI sidecar, the Piiranha transformer model (opt-in), JSONL feedback and human-approved pattern promotion.
Ties the PASS / LOG / REDACT / BLOCK actions to a weighted sum of format, algorithmic, BIN and context signals.
The Tamga red-team benchmark runs against Tamga's own scanner and publishes overall and per-category precision/recall/F1 plus P50/P95/P99/max scan latencies as JSON.
# from the proxy root make redteam-report # tamga/docs/benchmarks/redteam_latest.json
The full edition (36 pages, architecture diagrams and latency budget details) is shared only on enterprise request. Write to the sales team and we will send it under NDA within 24 hours.
