SASI middleware promises 15–50ms safety hooks
A new SASI SDK was introduced as middleware for deterministic LLM compliance — it intercepts inputs for obfuscation, PII redaction and attack filtering in 15–50ms to avoid slow LLM‑as‑judge delays introduced. Fast, local interception like this reduces latency and gives platform teams a predictable place to attach compliance logic.
TechViz brands the product as SASI — "Safety & Symbolic Intelligence" — and describes it as a pre‑LLM middleware SDK designed to enforce deterministic policy checks before model invocation. (techviz.us) The vendor advertises stateful "Risk Momentum" that tracks conversation trajectory across multiple dimensions to detect psychological volatility and boundary erosion, not just per‑message checks. (techviz.us) SASI includes a "Fail‑Closed Liability Protection" mode and a "Hard Safety Layer" that the site says will forcibly redact PII and trigger standardized crisis responses when anomalies are detected. (techviz.us) The public pages list cryptographic, tamper‑evident receipts and "Regulator‑Readable Output" aimed at auditability for FDA 524B, EU AI Act reviews and errors‑and‑omissions defense. (techviz.us) TechViz cites early integrations with apps named MyTrusted.ai, Bibbit.ai and a mental‑health assistant called Mentra, and a SASI demo video is available on the company YouTube channel. (techviz.us) Positioned as a pre‑LLM enforcement layer, SASI is presented as a deterministic choke point for compliance and forensic logging, implying platform teams would route agent inputs through SASI as a sidecar or proxy for enforcement and auditable receipts. (techviz.us)
