Platform · Intelligence

Sentinel

Agents that handle exceptions and earn autonomy.

An event-driven framework. Business transactions raise events; Sentinel recognises what kind of incident each is, decides whether to resolve it automatically, ask a model, or hand it to a person, records the outcome, and learns from it.

The framework is domain-free. Each product adds an overlay: a registry of exception patterns plus its own actions, read-only tools and receivers.

Recognise and route

Signed webhooks, polled APIs and database tables become fingerprinted cases.

Resolve safely

Playbooks, models with read-only tools, and people, all through guarded, idempotent actions.

Learn with sign-off

Trust is promoted per pattern only when a person approves.

Run without code

A console edits exceptions, people, roles and connectors as versioned configuration.

THE PIPELINE

From event to learning.

01 · Event02 · Fingerprint03 · Group & inhibit04 · Registry match05 · Route: auto · model · person06 · Guarded action07 · Disposition08 · Learning
WHAT IS GUARANTEED

Autonomy with a ceiling.

No event processed twice

An inbox table with a unique dedupe key per event.

No side effect executed twice

Every action step claims an idempotency key; re-drives skip completed steps.

Irreversible actions need a person

Only reversible or safe actions run unattended.

The model can't go rogue

Allowed actions only, schema-checked parameters, read-only tools, a step cap, then a person.

Trust is earned slowly, lost instantly

Promotion needs a track record and a person's sign-off; any reversal demotes at once.

Personal data stays out of models

Masked before every model call and restored only for the action.

Full audit

Every decision, attempt, compensation and reversal goes to an append-only log.

Blast radius capped

Per-pattern circuit breakers and kill switches.

OPEN-SOURCE STACK

Self-hosts anywhere.

Everything runs on open-source components. The frontier model is optional.

System of record, queue, learningPostgreSQL with pgvector
AnalyticsDuckDB, read-only
Frontier modelClaude (optional)
Small model + embeddingsOllama on a shared host
API, workers, consoleFastAPI, Python, plain HTML/JS
Sign-inTokens or any OpenID Connect provider
MonitoringPrometheus + Grafana
PackagingDocker Compose with nightly backups
Next step

Build on a foundation you can trust.

Tell us about your domain and we will show you how the platform fits.

Talk to us →