Inspiration
Modern payment operations teams face a daunting challenge: financial fraud rings operate at machine speed, yet automated systems that trigger unilateral actions (like refunds or account freezes) create massive compliance and financial risks. Inspired by the need for transparent, explainable, and human-controlled AI, we built AgentShield.
What It Does
AgentShield is a cloud-agnostic, autonomous multi-agent payment fraud investigation and prevention platform. It bridges automated AI threat detection with human-governed financial remediation. Key capabilities include:
- Multi-Agent Investigation Pipeline: Six specialized autonomous agents sequentially analyze transaction metadata, verify identities, uncover graph-based fraud clusters, synthesize forensic findings, evaluate risk, and recommend guarded remediation actions.
- Universal Payment Gateway Adapter: A cloud-agnostic payment gateway layer (
backend/app/gateway/) supporting live payment sandboxes (defaulting to OpenAPI/REST gateway endpoints) alongside a zero-risk local simulation mode. - Pluggable LLM Engine: Seamlessly switches between Amazon Bedrock (Claude), Google Gemini, OpenAI-compatible endpoints, and a zero-config
DETERMINISTIC_ENGINEfallback. - Forensic Incident Export: One-click generation of structured JSON and Markdown forensic dossiers for compliance and audit reviews.
How We Built It
- Backend: Python, FastAPI, SQLAlchemy, NetworkX, Scikit-Learn (Isolation Forest), Pydantic v2.
- Frontend: Next.js (App Router), React, TypeScript, Tailwind CSS, React Flow (Fraud Network Explorer).
- AI & Developer Tooling: Developed and steered using Kiro (with structured
.kiro/steering/and.kiro/specs/artifacts), supporting pluggable multi-LLM routing via Amazon Bedrock, Google Gemini, and OpenAI endpoints.
Challenges We Ran Into
Balancing fully autonomous multi-agent reasoning with strict financial safety was our biggest hurdle. We solved this by enforcing a Human-in-the-Loop Approval Gate—autonomous agents never execute unilateral refunds or reversals; every financial operation requires explicit human approval, capture-state verification, and immutable audit logging.
What We Learned
We learned that the future of enterprise AI isn't black-box automation, but explainable collaboration—giving analysts clear evidence dossiers, confidence scores, and auditable control gates.
Built With
- and
- aws-bedrock
- fastapi
- google-gemini
- networkx
- next.js
- openai
- tailwind-css
- typescript
- ython