X-RAIL

Explainable Risk Assessment & Insights Loop

Comment

X-RAIL

A multi-agent insurance AI platform built with Google Cloud + ADK

🧠 Inspiration

Insurance actuary and underwriting often hides behind black‑box models. Underwriters and customers alike deserve transparency — especially when risk decisions affect customer pricing, future claims, and trust. The global insurance industry wrote approximately $7 trillion in premiums in 2023, and with growing regulatory pressure for AI transparency, the need for explainable solutions has never been greater.

Inspired by the power of multi-agent AI orchestration using Google’s Agent Development Kit (ADK), we built X-RAIL: a fully explainable, modular, and interactive risk analytics engine for insurers.

X-RAIL stands for:
Xplainable Risk Assessment & Insights Loop — reflecting a platform where data, models, explainability, and simulations form a closed, intelligent decision-making system.

But beyond insight, X-RAIL delivers automation. It simulates how real-world insurers operate: cross-functional teams pulling queries, building models, explaining risks, generating reports, and running scenario tests. We transformed this traditionally manual, siloed pipeline into an intelligent, coordinated system of agents — each focused, explainable, and orchestrated — to automate a complex business process end-to-end.

Along the way, X-RAIL also delivers powerful data analysis and actionable insights, surfacing patterns, risk factors, and predictive indicators through transparent dashboards and scenario simulations.

🔧 What It Does

X-RAIL is a multi-agent autonomous system that can process incoming insurance claims — whether structured or unstructured — to compute risk scores, generate regulatory-aligned explanations, and simulate "what-if" scenarios in real-time.

X-RAIL can streamline underwriting workflows, reduce manual review time by 60- 70%, and produce transparent, auditable outputs for every prediction.

🧱 Agent Architecture

  • 🧠 Conductor Agent (ADK Root Agent)
    Routes tasks to specialized agents based on input type and user intent. Manages session context and workflow loops with persistent state management.

  • 🔢 Risk Agent
    Predicts a calibrated risk score using self-trained XGBoost with isotonic calibration and returns confidence scores. Hosted on Vertex AI and fully SHAP-explainable and auditable.

  • 🔍 Explainability Agent
    Uses SHAP to provide regulatory-aligned insight into the top features contributing to a user’s risk score. Includes both audit trail and plain-language summary options.

  • 📊 Dashboard Agent
    Renders metadata facilitating Streamlit-based dashboards showing risk gauges, SHAP plots, simulation deltas, and compliance metadata.

  • 🔁 Impact Simulator Agent
    Supports "what-if" edits (e.g., income change, car type) to see updated risk scores, premium impact calculations, and automated recommendations. Includes multi-scenario comparison capabilities.

  • 📈 BQML Agent
    Trains and evaluates BigQuery ML models on user request with RAG integration for documentation. Supports feature selection, model comparison, threshold tuning, and multiple model types.

  • 🧮 BigQuery Agent
    Converts natural language queries into SQL using ChaseSQL and retrieves real-time claim history, location risk, or policy info with comprehensive error handling.

  • 📈 Analytics Agent
    Runs Python-based analysis (via Vertex AI Code Executor) to provide data summaries, charts, statistical analysis, correlation detection, and policy insights.

🌟 Key Features

Capability Description
✅ Risk Scoring Calibrated risk predictions with confidence intervals
✅ Explainability SHAP-based explanations of the scoring with regulatory compliance
✅ What-If Analysis Interactive scenario simulation and impact assessment
✅ Dashboard Streamlit-based risk visualizations and plots
✅ PDF Reports Auto-generated, branded risk assessment reports
✅ NL-to-SQL Natural language to BigQuery queries for insurance data
✅ NL-to-Analytics Natural language to Python based analytics for claims, policies and customer analysis
✅ BQML Training Custom BigQuery ML model training and evaluation
✅ Comprehensive Audit Input hashing, audit IDs, conversation history, and complete decision traceability

🏗️ How We Built It

Architecture

Architecture Diagram

⚙️ Tech Stack

🧠 Core Orchestration & AI

  • Google ADK (Agent Development Kit) – Multi-agent system orchestration
  • Vertex AI – Model hosting and online inference
  • Gemini 2.0 Flash – LLM for task routing and natural language handling
  • Vertex AI RAG Engine – Retrieval-Augmented Generation for querying BQML docs

📊 Data, Deployment & Storage

  • BigQuery – Scalable data warehouse for structured insurance data
  • Google Cloud Storage – Storage for SHAP plots, PDFs, and model artifacts
  • BigQuery ML (BQML) – On-demand ML model training and evaluation
  • Docker - OS-level virtualization to deliver software in packages called containers
  • Cloud Run - Fully managed platform that enables you to run your code

🤖 Modeling & Explainability

  • XGBoost – Risk scoring via gradient boosting
  • Scikit-learn – Preprocessing and model calibration
  • SHAP – Feature attribution and model interpretability
  • Matplotlib – Visualization of SHAP outputs

📋 UI & Reporting

  • Streamlit – Real-time dashboards and simulation interface
  • FPDF – Automated PDF report generation with risk + compliance info

🧮 Data & Development

  • Pandas, NumPy – Data transformation and numerical analysis with statistical validation
  • ChaseSQL – Converts natural language to optimized SQL with error handling
  • Pydantic – Data validation across agents with comprehensive input sanitization
  • Poetry – Dependency and environment management with isolated environments

🧪 Testing & Debugging

  • Pytest, Pytest-asyncio – Unit and async testing for agents
  • IPython, IPykernel – Notebook development and debugging with stateful execution

🔒 Security & Compliance

  • Google Cloud IAM – Role-based access control with granular permissions
  • Audit Logging – Decision traceability for regulators with input hashing and audit IDs
  • EU AI Act Ready – Aligned with transparency and explainability standards
  • Input Validation – Comprehensive input sanitization and error handling

🔄 Agent Interactions

  • LLM-Powered Routing: Gemini handles user input classification and workflow control with context awareness.
  • ToolContext Sharing: All agents share input, risk output, explanation, and dashboard state via ADK context with persistent session management.
  • Cloud Integration: Agents communicate with BigQuery, Vertex AI, and GCS seamlessly for real-time data and compute.

🚧 Challenges We Faced

🔄 Multi-Agent Coordination

Designing and orchestrating interactions between independent agents while preserving shared memory (tool_context.state) was both conceptually and technically complex. We solved this by implementing a state management system using ADK's callback context and designing clear agent interfaces with standardized input/output formats.

📊 Data Availability & Generation

Insurance datasets are typically private, regulated, and hard to access. To overcome this, we built a Python-based data generation engine that created three domain-specific datasets:

  • 100,000 synthetic customer records
  • ~50,000 synthetic claims
  • 100,000 synthetic policy entries

These datasets follow realistic statistical distributions (e.g., Gaussian for income, categorical for coverage type) and include controlled noise and outliers to mimic real-world behavior as close as possible.

🌐 Deployment Complexity

Deploying a distributed, dependency-heavy agent system to Cloud Run was non-trivial. Sparse documentation, complex image builds, and inter-agent dependencies made deployment debugging time-intensive. We ultimately containerized agents using Docker and configured environment isolation through Poetry.

🔍 Explainability Integration

Translating SHAP outputs into clear, human-readable insight — while maintaining regulatory standards — required model calibration, feature pruning, and interpretability-first design.

🔐 Cloud Service Integration

Handling authentication and seamless operation across BigQuery, Vertex AI, and ADK required careful permissioning and modular design.

🛡️ Advanced Simulation Complexity

Developing core idea for Implementing multi-scenario comparison with premium impact calculations required sophisticated state management, error handling for edge cases, and automated recommendation generation based on simulation results.

🚀 Accomplishments

  • ✅ Delivered a full-stack insurance analytics engine with modular, explainable agents and comprehensive audit trails.
  • ✅ Successfully orchestrated complex agent interactions via ADK’s multi-agent architecture.
  • ✅ Enabled both expert (underwriter) and non-technical (customer) users through NL input, visual dashboards, and downloadable reports.
  • ✅ Created an extensible platform — from solo claims to portfolio-level simulation.

🧠 What We Learned

  • Multi-agent systems demand thoughtful task decomposition and agent granularity with robust state management.
  • Explainability isn't a nice-to-have — it's foundational in regulated industries and requires comprehensive audit trails.
  • Google's ADK provides a powerful backbone for LLM-driven orchestration and context-aware decisioning with persistent state.
  • Designing for auditability and transparency unlocks business trust and adoption while ensuring regulatory compliance.
  • Robust error handling and fallback mechanisms are critical for production-grade multi-agent systems.

🔮 What’s Next

  • 🛡️ Fraud Detection Agent – anomaly detection on claim patterns
  • 📱 Mobile App – for field agents and mobile-first customers
  • 🧾 Regulatory Compliance Agent – maps decisions to legal requirements
  • 📦 Portfolio Simulator – simulate multiple claims, assess book-wide risk
  • 🧠 Counterfactual Reasoning – “How could this risk score have improved?”
  • 🚗 Telematics Integration – Incorporate real-time vehicle data for dynamic, usage-based risk assessment.
  • 📊 Advanced Analytics – Enhanced statistical modeling and predictive analytics capabilities

💡 Why X-RAIL Stands Out

“A claim comes in. Multiple AI agents collaborate, reason, explain, and simulate outcomes — all within seconds. The underwriter sees a full dashboard. The customer gets an understandable explanation. And regulators see a compliant audit trail.”

  • Multi-agent intelligence: orchestrated via ADK with persistent state management
  • Fully explainable: SHAP, NLP, and visual insight with top 4 factors and directional impact
  • Real-time simulation: change inputs, re-score live with premium impact calculations
  • Built on Google Cloud: scalable, secure, production-ready
  • Comprehensive audit trails: input hashing, audit IDs, and complete decision traceability

🎯 Real-World Impact

X-RAIL helps insurers:

  • Reduce manual review time by 60–70%
  • Build customer trust through transparency
  • Satisfy compliance laws like the EU AI Act
  • Support underwriters with real-time simulations, premium impact calculations, and recommendations.

Built With

Share this project:

Updates