Property Flood Claims Triage Agent with CrewAI and Pydantic AI Structured Validation
During intense precipitation surges caused by El Niño, property insurance claims spike by orders of magnitude. Adjusters face immense backlogs separating minor baseline claims from critical structurally compromised properties requiring immediate field inspection and emergency response. In this challenge, you will orchestrate a collaborative multi-agent claims triage network using CrewAI validated by Pydantic AI structured models. The team will consist of an Image Analysis Specialist, a Policy Contract Inspector, and an Emergency Prioritization Lead working together. Your solution will ingest claim notes, structural damage drone photos, and property flood zone overlays, producing typed Pydantic models with actionable priority rankings and automated settlement recommendations.
What you are building
The core problem, expected build, and operating context for this challenge.
Build a multi-agent claims triage system using CrewAI and Pydantic AI to validate and prioritize property flood damage claims.
How work is evaluated
Evaluates the accuracy of priority classification and validity of output Pydantic data models.
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.
pydantic_schema_check
Output passes Pydantic model validation without validation errors
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
triage_priority_accuracy
Accuracy of priority level assignment (URGENT, STANDARD, LOW) • target: 0.95 • 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
Orchestrate collaborative role-playing agents using CrewAI.
Enforce strictly validated output structures using Pydantic AI schemas.
Extract damage indicators from unstructured claim notes and multimodal photo feeds.
Automate claims routing into fast-track, standard, or fraud investigation queues.
Reference links and supporting material
Synthetic database of property flood claim text logs, water depth readings, and visual asset summaries.
Challenge setup
Evaluates the accuracy of priority classification and validity of output Pydantic data models.
Challenge input
JSON with property_id, water_depth_cm, photo_summary, claim_notes
CrewAI
Mandatory framework for agent roles and collaboration.
Pydantic AI
Ensures structured output data adheres to schema.
Evaluated output
JSON matching ClaimTriageResult schema (priority, fast_track, estimated_loss)
- Output passes Pydantic model validation without validation errors
- Accuracy of priority level assignment (URGENT, STANDARD, LOW) • target: 0.95 • range: 0-1
- Benchmark: ClaimsTriageBench
- Triage Priority Accuracy target: 0.95
- 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
- CrewAI · Required
- Pydantic AI · Optional
- Aim · Optional
- Pydantic AI · Optional
Evaluation contract
- pydantic_schema_check · Weight 1
- triage_priority_accuracy · Weight 1
Metrics
- triage_priority_accuracy · Target 0.95
Benchmarks
- ClaimsTriageBench
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON with property_id, water_depth_cm, photo_summary, claim_notes
Output: JSON matching ClaimTriageResult schema (priority, fast_track, estimated_loss)
Execution harness: Python
Evaluation tools
- CrewAI · Multi-Agent System
- Pydantic AI · Type Validation
Recipe freeze marker
No recipe freeze time is recorded.