Private companion demo

Fraud detection and investigation platform

A company-neutral companion page for an AI-native fraud investigation platform. The platform proposes, but the system of record decides.

AgentCore Bedrock SageMaker LangFuse Kafka Human approval

Architecture visual for the fraud detection companion demo

Core thesis

Augment investigation, not decisioning

The AI layer gathers evidence, summarizes findings, and recommends next steps. Sensitive actions stay policy-gated and human-approved.

• typed internal tools over free-form agent behavior
• retrieval + tool use first, selective fine-tuning second
• unsupported claims are blocked or downgraded
• observability and evaluation are part of the product

System shape

Six-layer architecture

1. Analyst experience

2. Agent orchestration

3. Typed tools

4. Knowledge + retrieval

5. Bedrock + SageMaker model layer

6. Controls: LangFuse, MLflow, CloudWatch, IAM

Interactive mock demo

Select a case and inspect the investigation output

This uses mocked data from the prototype to show how an analyst-facing experience could present facts, model signals, recommendation quality, and approval requirements.

Facts

Model signals

Recommendation

Confidence

Approval

Tools used