Banking AnalyticsLive on BigQuery

Space Financial

A regional bank analytics platform built end-to-end: synthetic data generation, a BigQuery data warehouse modeled with dbt, an interactive BI dashboard, and an AI analyst that writes SQL and renders charts on demand.

Live Dashboard →AI Analyst (invite only)GitHub

2.37M

Data Rows

dbt Models

Mart Tables

AI Agent Nodes

Architecture

Data Generation

Python + Faker2.37M rows across 15 raw tables, 3 years of history

Data Warehouse

Google BigQueryraw → staging → marts with 16 mart tables

Transformations

dbt Cloud15 staging views + 16 mart models, SCD2 customers

AI Analyst

LangGraph + Claude7-node agent: intent → SQL → validate → execute → chart → narrative

BI Dashboard

Next.js + RechartsLive BigQuery queries, filterable charts, no caching delay

Auth Gate

NextAuth v5 + bcryptInvite-only credentials, signed JWT sessions

Key Engineering Decisions

SCD2 for customer history

dim_customer tracks changes over time with valid_from/valid_to and is_current. Queries use is_current=true for current state, preserving full audit trail.

Safety layer on AI SQL

Every query from Claude passes keyword blocking (no INSERT/DELETE/DROP), table allowlisting, and a 15-second timeout before hitting BigQuery. The AI cannot mutate data.

Native BigQuery surrogate keys

Surrogate keys use to_hex(md5(...)) natively in BigQuery — no dbt_utils dependency required, avoiding package import errors.

CAST AS FLOAT64 for NUMERICs

BigQuery NUMERIC type serializes as strings in JSON. All aggregations cast to FLOAT64 to ensure correct numeric parsing in the Next.js API layer.