Prompts and outputs. Classified, gated, logged.
The Engine sits between your application and your AI providers, scoring each message against your workspace policy in real time, blocking the risky ones, queueing the borderline ones for human review, and writing each decision to an append-only audit log.
Risk, classified and visually rendered.
Pulse is the live web of everything moving through the Engine. Each particle is a classified message; severity is colour-coded; reviewer corrections fire visible lightning bolts so operators can see the system learning. Click any particle for the full payload, verdict reasoning, and reviewer history.
- Severity-coloured particles · drag, hover, click to inspect
- Same-category mesh links surface emerging risk clusters
- Reclassification flashes when reviewers override the model
- Long-press focus mode dims everything except a cluster
Four hops, every message.
One URL change in your client, or one click to install the browser extension. Each request passes through the same four-stage pipeline before it ever reaches your AI provider.
- 01Intercept
Egress proxy at the SDK or REST layer, or the MV3 browser extension for ChatGPT, Claude, Gemini, and Copilot.
› path: client → engine - 02Classify
Scored against your workspace policy by the hosted classifier or your own fine-tuned model.
› verdict: severity + category + reasoning - 03Enforce
High severity blocked at the boundary. Medium queued for human review. Low passes through with metadata attached.
› action: block · review · pass - 04Audit
Each decision is written to an append-only ledger. Actor, timestamp, rule version, and the model's reasoning are all preserved.
› delivery: webhook · SIEM · ledger
What you get on day one.
Rule-based + LLM-classifier verdicts. Per-workspace policies. Severity thresholds you control.
A reviewer-friendly inbox for medium-severity events. Approvals, overrides, and comments propagate to training data.
Every override is a labelled training sample. Your classifier gets sharper without manual data engineering.
MV3 for Chrome / Edge / Brave. Governs prompts to ChatGPT, Claude, Gemini, and Copilot before they leave the browser.
Per-decision ledger. JSON webhook delivery + Slack bot for high-severity events. Retention configurable per plan.
Plug in OpenAI-compatible endpoints (vLLM, SGLang, custom). Run classification on your own infrastructure.
The extension surfaces which AI tools your team is already using across a curated catalogue, with a per-tool regulatory dossier and workspace risk register.
Stream every governance and discovery event to Splunk, Microsoft Sentinel, Elastic, or any CEF/LEEF receiver. Encrypted credentials, delivery health, auto-pause on failure.
Built for teams who need to ship AI safely.
Drop the Engine in front of OpenAI / Anthropic / Bedrock with a one-line URL change. Keep your existing SDK, get policy enforcement + audit trail for free.
Designed to produce evidence aligned with SOC 2, ISO 27001, FCA SYSC 4, and NIS2. Each decision is keyed by user, workspace, timestamp, and rule version.
Self-hosted in your VPC for data residency. SSO, SCIM, DPA, and MSA available. Customer success contact assigned on Enterprise engagements.
Free during the pilot. Enterprise when you scale.
Kirtonic is in early access. The platform is free during the pilot and includes the full Solo feature set. Higher classification volumes, SSO, SCIM, and self-hosted or managed deployment are scoped under Enterprise.
Solo tier features, reviewed by the team.
Self-hosted in your VPC, or managed by Kirtonic.
Integrate the Engine. Build your audit trail.
Early-access accounts are free during the pilot and include the full Solo feature set, so you can integrate the SDK, run real traffic through the Engine, and review the audit output before any commercial discussion.