FraudSentinel

Inspiration

Card-not-present fraud costs the industry billions, yet most systems just output a risk score and move on. We wanted an agent that investigates like a human analyst — and actually controls whether the money moves, before the payment clears.

What it does

FraudSentinel is an autonomous fraud-investigation agent that sits inside the payment flow.

A real payment — Card, Apple Pay, Google Pay, or PayPal — is authorized through Stripe but held: no funds are captured yet. The agent then investigates with a full toolchain — account history, an ML risk scorer, compliance rules, a card BIN pre-check, a threat feed, and an IP / network policy that refuses datacenter and VPN traffic.

Gemini reasons over the dossier and decides ALLOW, FLAG, or BLOCK. The verdict moves the money: ALLOW/FLAG captures the funds, BLOCK voids the authorization — and on a block it drafts a Suspicious Activity Report.

Analysts can ask a streaming chatbot to explain any decision in plain language, review a case queue with human-in-the-loop override, and watch analytics and drift. Every tool call, language-model reasoning step, evaluator, and SAR draft is exported to Arize as OpenInference traces — fully auditable.

How we built it

• Google Gemini 2.5 Flash for agentic reasoning and function calling
• FastAPI backend with WebSocket streaming and SSE
• React + Vite + Tailwind analyst console (light/dark, GSAP motion)
• Stripe manual-capture PaymentIntents for agent-gated card and wallet payments
• scikit-learn RandomForest risk scorer
• Arize OpenInference / OpenTelemetry tracing at every step
• Deployed with Docker + nginx + Cloudflare + Let's Encrypt on a VPS

Challenges we ran into

Making an agent actually gate real money — authorize, hold, then capture or void — through Stripe; routing Apple Pay, Google Pay and PayPal through the same agent; building theme-aware Stripe Elements; and getting reliable Arize span export at every reasoning step.

Accomplishments that we're proud of

A real fraud-control loop, not just classification. Four payment methods, all agent-gated. Full observability in Arize with 900+ spans. Live, deployed, and secured behind a single demo account.

What we learned

Manual-capture authorization is the primitive that lets an agent gate payments. OpenInference tracing makes agent reasoning auditable enough for production fraud operations.

What's next for FraudSentinel

Real Apple Pay merchant certification, more payment processors, a fine-tuned risk model, and Arize evaluators and monitors running in production.

Built With

• arize
• cloudflare
• docker
• fastapi
• gemini
• javascript
• nginx
• openinference
• opentelemetry
• playwright
• python
• react
• scikit-learn
• stripe
• tailwind-css
• vertex-ai
• vite
• websockets