Ledgerline: Governed Retail Banking Data Q&A

A retail-banking analytics demo built on synthetic data. A LangGraph agent answers questions only through a governed semantic layer, which compiles approved aggregates to read-only SQL and returns metric definitions, the executed query, and an audit trace.

Proof points

  • 8 Certified Metrics
  • 173 Automated Tests
  • Synthetic Data ยท Read-Only Queries

Technologies

Python, FastAPI, LangGraph, Next.js, SQLite, Semantic Layer, Vercel

Links

Problem

Business users need understandable answers from banking data without granting an LLM unrestricted SQL access or exposing customer-level rows.

Approach

Defined a read-only warehouse and a governed semantic layer for eight metrics, then connected a LangGraph workflow to validate, execute, and audit permitted aggregate queries. The web demo also compares the live workflow with a deterministic rule-based fallback.

Results

Deployed a working demo against synthetic data with metric definitions, SQL, traces, request limits, and a documented failure path. The evaluation is about system behavior; it does not establish outcomes for a real bank.

Ledgerline: Governed Retail Banking Data Q&A

Ledgerline combines a synthetic retail-banking warehouse, semantic model, LangGraph question-answering workflow, FastAPI service, and dashboard. Eight certified metrics define their formulas, grain, additivity, filters, and synonyms. The agent cannot write free-form SQL, retrieve customer-level rows, or modify data; it must use the governed query path and show the definition, SQL, and trace.

The public demo uses synthetic data and per-visitor limits. The source repository is private. The README documents 173 automated tests; these establish software behavior, not the business effectiveness of the analytics.