Agents for financial audit.

What the pipeline touches

Three connector kinds in one run is the point. It is the difference between reconciling genuinely distributed data and demonstrating against a single database.

• A Postgres ledger, queried through the proxy under whatever row filter the requester's grant allows.
• Receipt documents in object storage, reached under a key prefix rather than an open bucket.
• An external rate API, reached over HTTP under a URL allowlist declared in the definition.

How the report gets built

An agent task gathers the permitted data and emits a structured report specification - sections, prose, dataset references, chart definitions. That output is a Record, and it is validated against its registered schema before anything downstream sees it.

A deterministic render task then composes the finished site from a hand-built template and component library held as a Resource. The presentation comes from code polished once, not from a model improvising markup on the day of the meeting.

The artifact

The output is a File: a self-contained page with the data snapshot baked in, so charts render and filter in the browser with nothing running behind them. It opens like a web page because it is one.

An audit trail built for finance

Every query, every document fetched, every scope decision, and every repair attempt is an event in the run log. When a controller or an auditor asks where a figure came from, the answer is a recorded sequence rather than a reconstruction from memory.

Runs where your data has to live

Deploy in our managed cloud, in a private VPC inside your own cloud account, or on-premises with no outbound calls. Financial data can stay inside your network boundary.