Google ADK Fraud Guard: Multimodal Hurricane Claim Inspection
Post-hurricane claims processing suffers from fraudulent estimates and mismatched damage evidence across property adjusters. Build a Google ADK multimodal agent integrated with GPT-5 Pro to cross-evaluate satellite imagery, field reports, and repair invoices to achieve an F1 score above 0.88 on fraud detection.
What you are building
The core problem, expected build, and operating context for this challenge.
Implement a multimodal Google ADK agent that cross-references pre/post storm imagery with adjuster notes and invoice line items to identify fraudulent claim spikes.
How work is evaluated
Evaluates the agent's ability to classify fraudulent claims correctly based on satellite images, loss descriptions, and invoice line items.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
json_schema_compliance
Validates that the output matches the required schema.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
f1_score
Harmonic mean of precision and recall for fraud classification. • target: 0.88 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Build a multimodal agent using Google ADK and GPT-5 Pro for insurance fraud detection.
Integrate geospatial satellite imagery parsing with unstructured claims text.
Implement automated discrepancy analysis between field adjuster estimates and image ground truth.
Achieve an F1-score of at least 0.88 on post-catastrophe claims validation.
Reference links and supporting material
Synthetic post-hurricane claims dataset containing 500 property records with paired aerial imagery metadata, adjuster narrative notes, and billing records.
Challenge setup
Evaluates the agent's ability to classify fraudulent claims correctly based on satellite images, loss descriptions, and invoice line items.
Challenge input
JSON containing claim_id, image_url, adjuster_report_text, line_items
Google Agent Development Kit
Primary framework for orchestrating multimodal AI agent tasks.
GPT-5 Pro
High-capacity multimodal LLM for complex legal and spatial reasoning.
Evaluated output
JSON containing risk_score, fraud_detected (boolean), reasoning_summary
- Validates that the output matches the required schema.
- Harmonic mean of precision and recall for fraud classification. • target: 0.88 • range: 0-1
- Benchmark: Multimodal Fraud Evaluation Suite
- F1 Score target: 0.88
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Modal · Optional
- Google · Required
- Aim · Optional
- Modal · Optional
Evaluation contract
- json_schema_compliance · Weight 1
- f1_score · Weight 1
Metrics
- f1_score · Target 0.88
Benchmarks
- Multimodal Fraud Evaluation Suite
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing claim_id, image_url, adjuster_report_text, line_items
Output: JSON containing risk_score, fraud_detected (boolean), reasoning_summary
Execution harness: Python
Evaluation tools
- Google Agent Development Kit · Agent Framework
- GPT-5 Pro · LLM Engine
Recipe freeze marker
No recipe freeze time is recorded.