LoginSign Up
← Back to all sparks
S

Superwise

ANALYTICS
Velocity5.0

Machine learning model observability and monitoring platform

superwise.ai ↗

Superwise lets teams watch a guardrail fire before letting it block anything.

llm-gatewayguardrailsphi-detectionai-governancedashboardstoken-optimization
◆Current state
Superwise's Sentinel gateway is filling out its guardrail set and the reporting around it. Version 1.41 added a PHI detector that uses a model to find health details in free text and redacts or blocks them before the prompt reaches the LLM. Version 1.42 adds a Flag action that records a finding without changing the request, and dashboards now split traffic into passed versus findings, with blocked, remediated and flagged counts and custom ranges up to 180 days.
◆Where it's heading
After the token-saving work of September and October, the focus has shifted to guardrails that a compliance team can roll out gradually: generated test cases, then a log-only mode, then reporting that shows how often each rule would have acted. That lowers the cost of turning on stricter checks such as PHI redaction in production.
◆Prediction
Expect more context-aware detectors in the PHI mould, likely for other regulated data types, plus per-guardrail reports built on the new Flag counts.

◆Recent moves

  1. 1d ago

    Guardrails gain a log-only Flag mode; dashboards split outcomes

    Any guardrail can now Flag instead of block or rewrite, so teams can measure how often it would trigger before enforcing it. Dashboards count passed versus flagged, remediated and blocked traffic, which turns the guardrail work into numbers a compliance lead can review.

    View source ↗
  2. 9d ago

    PHI detection guardrail redacts health data before it hits the model

    A new guardrail reads context to catch patient names, conditions and health identifiers in prompts and redacts or blocks them. It is the first detector aimed at a regulated industry, extending guardrails beyond the format checks added in earlier releases.

    View source ↗
  3. 12d ago

    Generated guardrail test cases and guided client setup

    Each guardrail can now generate example test cases from its configuration, for text and image checks, and client setup is guided by Sentinel state. It extends the Test sandbox from 1.39 by removing the need to write the samples.

    View source ↗
  4. 19d ago

    Image sharpness guardrail and CLI view of a Sentinel's full policy

    An image sharpness guardrail joins image resolution, the language check gets more accurate, and CLI 0.4.2 shows a Sentinel's full policy. Steady expansion of the guardrail set, with admin tooling alongside.

    View source ↗
  5. 1mo ago

    Caveman output shortening, token-saving reports, new guardrails

    Caveman shortens model output at three intensities, and token savings now appear on dashboards, the home page and per interaction. It moves Sentinel from counting spend to reducing it, with language and image-resolution guardrails and a test sandbox in the same release.

    View source ↗
  6. 1mo ago

    Sentinel becomes the default product, with new Solo to Business plans

    Sentinel becomes the default product, guardrails become editable per Sentinel, Headroom compression arrives, and Solo, Pro and Business plans center limits on Sentinel usage. It sets up the token-optimization work that follows and shows where Superwise now puts its weight.

    View source ↗