Sentinal AI

Inspiration

Financial fraud detection systems can flag suspicious transactions, but investigating those alerts still requires significant manual effort. Compliance analysts often spend 6–8 hours collecting evidence, checking transaction histories, tracing account relationships, and preparing reports. We wanted to solve the gap between detecting fraud and actually investigating it. This inspired us to build Sentinel AI, a multi-agent AI platform that automates financial fraud investigation while keeping its findings explainable, evidence-based, and auditable. Our goal is to help financial institutions investigate suspicious transactions faster, reduce repetitive manual work, and give analysts the evidence they need to make informed decisions.

What it does

Sentinel AI uses five specialised AI agents to investigate suspicious financial transactions through a coordinated pipeline. Agent 1 — Anomaly Detector: Uses XGBoost and Isolation Forest to identify unusual transaction behaviour and highlight important risk indicators. Agent 2 — Evidence Gatherer: Collects and organises transaction history, customer information, device details, beneficiary information, sanctions checks, news mentions, and location-related evidence. Agent 3 — Network Analyser: Uses Neo4j graph analysis to trace money flows and identify suspicious patterns such as layering, multiple accounts sending money to a common beneficiary, and potential fraud networks. Agent 4 — Risk Assessor: Combines risk signals into a score and maps relevant findings to regulatory references and recommended actions. Agent 5 — Explanation Generator: Produces a structured investigation report using evidence gathered by the preceding agents. A rule-based validation layer checks the generated report against the available evidence before it is presented for review. A compliance dashboard brings together the case queue, risk scores, evidence, network visualisations, investigation reports, and analyst actions. The complete workflow is designed to reduce investigation time from hours to approximately 12 minutes, while maintaining traceability and human oversight

How we built it

We designed Sentinel AI as a modular, multi-agent system with a structured data pipeline. Python and FastAPI: Handle backend processing, transaction-alert ingestion, data validation, and API endpoints. LangGraph: Orchestrates the investigation workflow and manages the handoff of structured information between agents. XGBoost and Scikit-learn: Support supervised fraud detection, anomaly analysis, preprocessing, and model evaluation. Pandas and NumPy: Handle transaction data preparation and feature engineering. SHAP: Helps explain model predictions by identifying influential transaction features. Neo4j: Represents accounts and money transfers as a graph to analyse relationships and suspicious transaction networks. Claude API: Generates structured investigation reports from supplied evidence using strict instructions and predefined templates. Python rule-based validation: Checks report values, regulatory references, recommended actions, and source traceability. MongoDB Atlas: Stores case information, agent outputs, investigation reports, decisions, and audit records. React and Tailwind CSS: Power the compliance dashboard, with Recharts and React Force Graph supporting data and network visualisations. Our ML pipeline uses the IEEE-CIS Fraud Detection dataset, with PaySim and a custom synthetic Indian banking dataset providing additional testing scenarios. The architecture is modular, allowing individual agents and components to be developed, tested, and improved independently.

Challenges we ran into

One of our biggest challenges was ensuring that an AI-generated investigation report could be trusted in a compliance-sensitive environment. Language models can produce unsupported statements, so we designed a structured evidence-grounding approach in which the report generator uses supplied data rather than independently inventing explanations. Another challenge was analysing fraud that cannot be identified by examining a single transaction. We addressed this through graph-based network analysis, which helps uncover relationships between accounts and suspicious money-flow patterns. Financial transaction datasets also present challenges such as class imbalance, missing values, and complex feature relationships. We used complementary supervised and unsupervised models, feature engineering, evaluation metrics, and explainability techniques to address these issues. Finally, integrating multiple agents, validating their data handoffs, handling failures, and presenting complex findings in a clear dashboard required a modular architecture and careful testing.

Accomplishments that we're proud of

We are proud of designing Sentinel AI as an end-to-end investigation platform rather than just another fraud prediction model. Our key accomplishments include: Designing a five-agent architecture that covers anomaly detection, evidence gathering, network analysis, risk assessment, and report generation. Developing an ML pipeline with XGBoost, Isolation Forest, and SHAP-based explainability. Building a canonical transaction-data approach to support integration across different financial datasets. Designing graph-based analysis using Neo4j to investigate connected accounts and suspicious money flows. Creating an evidence-grounded reporting approach with rule-based validation and source traceability. Structuring the project into modular components with independent tests and clear development phases. Defining a compliance dashboard that brings investigation findings and analyst decisions into one place. Our vision is to make financial investigations faster, more consistent, and easier to audit without removing the human analyst from important decisions.

What we learned

Building Sentinel AI taught us that effective AI systems require more than accurate predictions. In financial investigations, a model must also explain its findings, provide traceable evidence, and support responsible decisions. We learned how supervised and unsupervised machine learning can complement one another, how SHAP can help explain predictions, and how graph databases reveal relationships that traditional transaction-by-transaction analysis may miss. We also gained experience with multi-agent orchestration, structured data contracts, API integration, model evaluation, automated testing, and full-stack application design. Most importantly, we learned to treat reliability, evidence validation, and human oversight as core parts of the architecture rather than features added at the end.

What's next for Sentinal AI

Our next steps are to strengthen the end-to-end integration of all five agents, improve evaluation across different transaction patterns, and expand testing with synthetic and benchmark datasets. We plan to improve the compliance dashboard, add richer interactive network investigations, enhance monitoring and error recovery, and make the system easier to deploy in different environments. We also aim to integrate authorised live financial data sources, strengthen security and access controls, and evaluate the reporting and risk-assessment workflow against realistic compliance scenarios. Our long-term vision is to make Sentinel AI a reliable AI-powered investigation assistant that helps financial institutions move from isolated fraud alerts to faster, evidence-backed investigations while keeping final decisions under appropriate human and regulatory oversight.