Luxury Commerce Fraud Inspection Agent using Claude Agents SDK and Replit Agent
High-value e-commerce returns require accurate authentication and rapid fraud detection to prevent financial leakage in premium retail. Build an automated return verification workflow using the Claude Agents SDK with extended thinking capabilities alongside Replit Agent tooling to inspect claim documentation and flag fraudulent buyer claims.
What you are building
Develop an automated luxury returns inspector using Anthropic's Claude Agents SDK with extended thinking to detect discrepancy signals in claim evidence.
How work is evaluated
Evaluates return fraud risk classification accuracy across sample claim scenarios.
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
serial_mismatch_detected
Checks whether serial number mismatches are identified correctly.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
fraud_detection_f1
F1 score on return fraud classification benchmark. • target: 0.92 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
Implement Anthropic's Claude Agents SDK with extended thinking enabled for deep reasoning.
Use Replit Agent tooling paradigms for rapid environment prototyping and service integration.
Build strict fraud evaluation criteria for high-value fashion return requests.
Resources
Return claims dataset containing item details, return descriptions, claim images metadata, and verified outcome labels.
Challenge setup
Evaluates return fraud risk classification accuracy across sample claim scenarios.
Challenge input
JSON containing original order item metadata, return claim reason, and claim photos metadata.
Claude Agents SDK
Extended thinking support for thorough audit reasoning
Replit Agent
Rapid agent environment setup and deployment
Evaluated output
JSON containing risk_score, decision, and detected_discrepancies.
- Checks whether serial number mismatches are identified correctly.
- F1 score on return fraud classification benchmark. • target: 0.92 • range: 0-1
- Benchmark: E-Commerce Fraud Detection Suite
- Fraud Detection F1 target: 0.92
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- Replit Agent · Required
- Replit Agent · Required
- Aim · Optional
- Deepgram · Optional
Evaluation contract
- serial_mismatch_detected · Weight 1
- fraud_detection_f1 · Weight 1
Metrics
- fraud_detection_f1 · Target 0.92
Benchmarks
- E-Commerce Fraud Detection Suite
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing original order item metadata, return claim reason, and claim photos metadata.
Output: JSON containing risk_score, decision, and detected_discrepancies.
Execution harness: Python
Evaluation tools
- Claude Agents SDK · agent_framework
- Replit Agent · ai_engineering_tooling
Recipe freeze marker
No recipe freeze time is recorded.