Sebastian
Request access
Platform · Governed finance data

Govern the finance data and approvals behind accounting AI.

Sebastian's governed finance data architecture separates shared identity from workspace-scoped finance records, preserves selected sensitive fields with workspace-specific encryption, versions close evidence, and persists agent execution across human approval interrupts.

Request accessInspect the close evidence model →
Source-bound evidence

A source version retains more than the rows.

Close evidence snapshots can retain the query parameters, control totals, row count, canonical content, and SHA-256 digest needed to identify the source used by a review. A later verification recomputes the digest on the server.

  • Versioned snapshots preserve historical review inputs
  • Parameters and control totals keep the extract scope inspectable
  • Server-side content verification detects mutation of stored snapshot data
sebastian · platform / source evidenceIllustrative re-enactment · demo data
Example close task illustrative files
illustrating…
  • csvexample-source.csvIllustrative task file · 24 KB
  • xlsxreview-workbook.xlsxIllustrative workbook · 38 KB
  • jsoncontrol-totals.jsonIllustrative snapshot · 4 KB
  • pdfreview-note.pdfIllustrative decision · 12 KB
Example files shown. Source snapshots and task attachments persist as separate records.
Durable agent work

An AI approval workflow can pause execution without losing state.

The finance agent data layer persists conversation and execution state outside the HTTP request. Designated write tools can interrupt for human approval and resume afterward, keeping the proposed input and resulting record in one durable workflow while supporting accounting data security.

  • Queued execution survives beyond one request-response cycle
  • Designated write tools pause before execution
  • Approved journal creation produces a pending-review draft rather than a direct post
sebastian · platform / approval interruptIllustrative re-enactment · demo data
Illustrative balanced journal draft
JE-DEMO-001 · example dataSebastian
Preparing example…
AccountDebitCredit
6000 · Example expense100.00
2100 · Example accrual100.00
Balanced100.00100.00
Illustrative optional reasoning trace
  1. 1Read persisted period-level journal totals
  2. 2Mapped the proposed debit and credit lines
  3. 3Validated that total debits equal total credits
Architecture facts

The implemented layer, stated without connector theater.

Workspace schemas

Each tenant receives a workspace-specific Postgres schema for its business data.

tenant boundary
Shared identity catalog

Global user identity and workspace membership live in a separate system catalog.

identity · membership
Private-record RLS

Row-level security applies to user-private records inside the workspace data model.

scoped access
Scoped field encryption

Agent messages, checkpoints, memory, filenames, and connector credentials use per-workspace AES-256-GCM keys.

selected fields
Versioned evidence

Close source snapshots retain provenance, control totals, row counts, and content verification.

verifiable
Connector boundary

A live NetSuite mirror, full-ledger sync, drift service, and ERP write-back are not represented as shipped capabilities.

implementation gate
Go deeper

See the work up close.

Close evidence

Inspect the source snapshot, content hash, and verification model in detail.

Explore

Journal entry automation

See how a designated agent write pauses for approval and enters pending review.

Explore

SHA-256 evidence protocol

Test content integrity with a worked canonical-payload and digest example.

Explore

Keep your ERP. Close and plan in one pane.

AI for accounting and FP&A on the ledger you already run. No migration. No rip and replace.

Request accessSign in
Agents at work · Matching an invoice to its PO line…