Skip to content

demo-script.md

Latest commit

 

History

History
60 lines (46 loc) · 3.27 KB

demo-script.md

File metadata and controls

60 lines (46 loc) · 3.27 KB
Raw

Demo script

The interview walkthrough. The automated end-to-end tests (tests/Reva.E2E) drive this exact flow in a real browser, so "it works" is provable on screen.

Run it

dotnet run --project src/Reva.Web/Reva.Web.csproj -- --seed-demo

Open the workspace. (Local runs don't auto-open a tab; the packaged build does.)

The demo corpus

Seeded on first run — one document per kind, chosen to show range:

Document Type Shows
orion-property-cat-xl-jan-2025.eml Premium bordereau (broker cover-note email + CSV attachment) The hero: extraction + reconciliation (the stated cover-note totals deliberately disagree with the attached line items) + source citations
technical-account-statement.txt Statement of account Clean extraction, zero exceptions, ready to approve
operations-note.txt Unknown Never hard-rejected — a non-reinsurance note still becomes a low-confidence reviewable record

What the engine processes today: Excel/CSV, PDF (digital + scanned via OCR), Word, PowerPoint, emails with attachments, plain text, and anything else best-effort. Document types it understands: premium & claims bordereaux, statements of account, slips, loss runs, claim notices. See docs/research/reinsurance-landscape.md for the domain grounding.

Walkthrough (about 5 minutes)

  1. Workspace — the dense operations view: KPI strip, a work queue, and the segmented filter (All / Needs review / Clean). Filter to Needs review to show triage.
  2. Import anything — drop a file (even a random one). It is never rejected; it becomes a reviewable record and opens straight into review.
  3. Review & adjust — the split view: the document on the left, extracted fields on the right. Hover a field and its value lights up in the document (the source citation). Edit a field; it shows as Reviewed once approved.
  4. Checks — the Detected-vs-Expected reconciliation cards, computed from the data, ranked by how badly they disagree.
  5. Export — pick a template from the Export menu (Bordereau line items, Lloyd's CRS 5.2, Canonical CSV, Full JSON) and download as Excel/CSV/JSON. Build your own on the Export templates page: choose columns, rename headers, pick the format, watch the live preview.

Why it is built well (talking points)

  • Customizable: Settings let you switch theme (light / dark / system), recolour the accent, rename the product, set confidence thresholds, pick a default export template, and manage data — all persisted and applied across the app.
  • Modular: parsing, OCR, classification, extraction, and reconciliation are separate, swappable pieces behind interfaces — add a parser or an LLM extractor without touching the UI.
  • Extensible: never hard-rejects; new document types and export templates slot in.
  • Trustworthy: confidence is computed (not faked), every value is traceable to its source, and corrections are audited.
  • Proven: real Playwright end-to-end tests run this whole demo on every change.

Architecture

  • Backend-first, contract-driven; the UI injects the workflow directly.
  • Native .NET parsers + offline PaddleOCR; EF Core (SQLite default, SQL Server by config).
  • Human-in-the-loop review for auditability. Synthetic data only.