The prototype is a functioning add-on, not a mockup — it runs live inside a Google Sheet, reconciles synthetic data, and drafts an actual AI-generated email.

This page walks the build and the evidence it works.

How it's wired

The add-on has two halves that talk to each other:

• Code.gs runs server-side in Apps Script. It builds the menu, reads the two data tabs, runs the deterministic recon engine, and makes the Gemini call.
• Sidebar.html is the panel the analyst sees. It requests results and renders the worklist.

They communicate through Apps Script's google.script.run bridge. The recon engine reads the sheet's ranges, matches on trade ID, computes variances, separates one-sided trades, and returns a ranked, classified result object. The Gemini API key is stored in Script Properties, never hardcoded — a deliberate security practice, and the right habit to show in a portfolio.

The synthetic dataset

Two tabs, Internal and Broker, of OTC-derivative variation-margin positions (interest-rate swaps, CDS, swaptions, FX forwards). The data was engineered to exercise every branch of the engine:

• Several trades that match within tolerance (noise that should be filtered out)
• A few with genuine valuation gaps over the CSA Minimum Transfer Amount (the value breaks)
• One trade on the broker file only and one on the internal file only (the population breaks a variance sort would miss)

The result

Running Reconcile produces the exact target output, which is how the engine is verified:

• Net dispute: +$327,500