- contact@verticalserve.com
- Every policy checked against what was agreed · part of the Insight broker suite
InsightProofing is deployed into the brokerage’s own environment. Client policies, documents and the decision record stay there; the model is called through your own model account.
A Python (FastAPI) API and a React workbench. Production refuses to start without a configured database — there is no fallback to a development instance.
One schema on the platform Postgres, versioned with Alembic migrations. The database password is a secret reference resolved at start, never in the URL.
Documents go to the object store when one is configured and never to local disk; otherwise only their text is kept.
Through your AWS account. Switch the agents or the copilot off and the engines’ findings stand and say so.
Sign-in through the Insight identity platform with roles: account executive, account manager, practice leader, compliance.
Account executives and managers see the checks they own or manage; practice leaders and compliance see every check. Outside scope is a 404.
Every row belongs to an organization; a token for an unknown organization sees nothing.
Every request is logged once with a correlation id; access tokens in a query string are redacted.
Agents and copilot answer in a fixed shape citing numbered facts; an answer with a number not in its facts is thrown away; prose markers are stripped before display.
Every action is capped at L2. What goes to a carrier or a client, and delivering a policy, is a person’s decision.
A trace per agent step with model, tokens, cost and time; a daily budget per agent, over which the fast model answers.
Prompts live in the pack and in InsightPrompts, overridable without a release; every run records the prompt it used.
The extraction, proofing and comparison engines are pure functions tested against generated documents with known errors.
Tests run in their own schema on the database and drop it at the end.
No test calls a real model: Bedrock is refused at the AWS client in the test run.