Tax anomaly detection with BigQuery ML and Gemma

In the public financial sector, the ability to detect fraud while keeping data strictly in-country can be a mission-critical requirement.

This solution provides a blueprint for addressing this challenge with Cloud de Confiance by S3NS. Agencies can deploy BigQuery ML on Cloud de Confiance by S3NS to analyze millions of tax declarations and identify high-risk anomalies in real time. Simultaneously, a local, open-weight Gemma RAG system uses vector search to map those anomalies directly to tax policies. This ensures context-aware compliance and full auditability, without making external API calls or moving data outside the sovereign boundary.

You can deploy this solution by following the accompanying reference implementation with Terraform.

Target audience

This solution is designed for national tax authorities or regulated financial agencies. It serves the following stakeholders:

  • Auditors & tax investigators who review real-time dashboards to investigate flagged anomalies and use the grounded Gemma AI assistant to generate traceable case summaries.
  • Compliance officers who centrally manage the system's legal knowledge base by uploading and indexing official tax policy documents within a secure, local interface.
  • Data scientists who securely access raw data using GKE-hosted Jupyter environments to iteratively train, validate, and redeploy anomaly detection models.

Core capabilities

  • Sovereign anomaly detection: Use BigQuery ML to screen massive datasets and flag high-risk anomalies instantly, ensuring data never leaves the sovereign boundary.
  • Retrieval-Augmented Generation (RAG): Run semantic vector searches across internal policy libraries to provide grounded, context-aware insights using the open-weight Gemma LLM.
  • Localized AI assistance: Generate summaries of policy violations through an AI assistant running entirely within the regional perimeter.
  • Dynamic knowledge management: Maintain an up-to-date compliance database by indexing official regulations as vector embeddings for immediate use in RAG pipelines.
  • Dynamic internationalization: Instantly switch UI language between English, French, and German without reloads or restarts.

Architecture

This architecture leverages BigQuery ML and a RAG pipeline with Gemma LLM to deliver secure, context-aware tax anomaly detection.

Tax Office

Components

The following technologies and Cloud de Confiance by S3NS services are used in this solution:

Component Tech Purpose
Data Python Generate tax data (TRAINING/NEW_FILING).
Model BigQuery ML Logistic Regression for anomaly detection.
LLM Gemma Open-weight model providing context-aware insights grounded in official tax policy.
RAG BigQuery Vector search for semantic matching and RAG grounding.
Container infrastructure GKE Container infrastructure for the tax office application and JupyterHub server.
Analysis Jupyter Notebook for ML model training and analysis/validation.
App Flask Web app for prediction visualization.

Reference implementation

A reference implementation of this solution with Terraform is provided in GitHub. Note that this is a proof-of-concept prototype built for demonstration purposes, and the implementation is not audited or secured for production use cases.

For prerequisites and deployment instructions, see Sovereign Tax Anomaly Detection with BigQuery ML and Gemma.