Sentinel

Inspiration

Financial fraud is no longer just about catching one unusually large transaction. Modern fraud is coordinated, adaptive, and often hidden across many accounts, devices, and behaviors. We wanted to build something that goes beyond simple rule-based detection and static thresholds. Sentinel was inspired by the idea that fraud should be understood as a live intelligence problem: not just “is this transaction suspicious,” but “does this fit a broader pattern of abnormal behavior, linked relationships, or emerging fraud tactics?” We wanted to create a platform that feels like a real fraud command center, where AI helps businesses detect what humans and traditional systems often miss.

What it does

Sentinel is a web-based fraud intelligence platform powered by Aegis, our multi-layer AI engine. It ingests financial events such as transfers, payments, logins, beneficiary changes, and account activity, then evaluates them in real time using multiple layers of fraud analysis. The system compares each event against account history, looks for suspicious relationship patterns across accounts and devices, detects unusual behavioral sequences, identifies novel anomalies, and applies deterministic fraud rules. These signals are fused into one final risk score and decision, such as approve, challenge, hold, or block. On the front end, Sentinel gives fraud analysts a live dashboard with real-time alerts, case management, network graph investigation, behavioral timelines, reporting, and an AI copilot chat that explains why an event was flagged and helps analysts investigate faster.

How we built it

We built Sentinel as a full-stack, production-oriented fraud platform. The frontend was designed as a modern web application for analysts and managers, with pages for live risk monitoring, alerts, cases, graph investigation, entity profiles, reports, and admin controls. The backend was built to manage event ingestion, alert creation, case workflows, audit logs, and scoring orchestration. At the core of the system is Aegis, our AI fraud engine. Instead of relying on a single model, we designed it as a layered system:

1. a tabular fraud model for structured transaction risk
2. a graph intelligence layer for linked-account and fraud-ring detection
3. a behavioral sequence layer for suspicious action patterns
4. an anomaly layer for unseen or emerging fraud
5. a rules engine for known fraud signatures and policy checks
6. a fusion layer that combines everything into one final decision We also designed the system so that real or replayed transaction data flows through the same internal event pipeline, which makes the platform extensible for both demo datasets and future production-style integrations.

Challenges we ran into

One of the biggest challenges was scope. Fraud detection at a realistic level is not just a model problem; it is also a product, data, systems, and operations problem. We had to balance ambition with what was realistically buildable. Another challenge was designing the platform so it felt production-grade without overcomplicating the first version. We also had to think carefully about how to make the AI powerful but still explainable, since fraud analysts need evidence, not just a black-box score. Data design was another major challenge. Fraud systems are only as good as the event schema, feature pipeline, and context they use. We had to think about how to structure incoming events, preserve account and relationship information, and make the system compatible with both replayed datasets and future live event ingestion. On the product side, it was challenging to make the frontend feel like a real investigation platform rather than just another dashboard.

Accomplishments that we're proud of

We are proud that Sentinel is not just a single fraud model or a static UI mockup. It is a complete fraud platform with a clear separation between the AI scoring engine, backend orchestration, and analyst-facing interface. We are especially proud of the layered design of Aegis, because it reflects how real-world fraud actually works: through transactions, behavior, relationships, and anomalies all at once. We are also proud of the product experience. The live risk dashboard, case workflow, network graph, and AI explanation flow make the system feel like something a real fraud operations team could use. Another accomplishment is that the platform was designed with extensibility in mind, so the same architecture can support demo replay now and more production-like ingestion later.

What we learned

We learned that strong AI systems are not just about choosing a powerful model. The real challenge is building the full loop: ingestion, context, scoring, explanation, workflow, and decision-making. We also learned that fraud detection works best when multiple perspectives are combined. Looking at only transaction size or only rules is not enough. Real fraud intelligence comes from combining structured signals, behavior, relationships, and anomaly detection. We also learned how important explainability is. An accurate fraud score is useful, but a score that an analyst can understand, trust, and act on is much more valuable. Finally, we learned that building a realistic AI product requires systems thinking: the frontend, backend, models, and data pipeline all have to work together.

What's next for Sentinel

The next step for Sentinel is to move from a replay-driven prototype into a more production-like streaming architecture. That means adding stronger real-time ingestion, richer online features, and more durable service-to-service infrastructure. On the AI side, we want to continue improving Aegis by strengthening the graph and sequence components, expanding anomaly detection, and improving score fusion and model monitoring. On the product side, we want to deepen the analyst workflow with stronger case collaboration, better reporting, and more advanced copilot capabilities. We also want to support more realistic integrations so Sentinel can evolve from a strong fraud intelligence demo into a scalable fraud operations platform. Ultimately, the goal is for Sentinel to become a system that helps organizations not only detect fraud faster, but understand it, investigate it, and adapt to it as threats evolve.

Built With

• docker
• fastapi
• openai
• python
• typescript