LedgerLens

Inspiration

Our approach drew primary inspiration from pattern-analysis algorithms and variable correlation techniques, such as Graph Neural Networks (GNNs). Furthermore, to counteract common AI biases and hallucinations, we implemented a three-agent architecture designed to ensure that the Large Language Model cannot fabricate data or make decisions without providing explicit justification. For data processing, we established a dual-methodology combining Bayesian networks with graph algorithms, significantly enhancing analytical reliability by requiring independent convergence toward similar conclusions.

What it does

The system ingests user-provided data and processes it using two distinct methods: a Bayesian network and graph algorithms. Once processed, the data is forwarded to the Investigator agent. The Investigator utilizes an Action Bank to select specific actions based on the hypothesis under investigation and the questions it needs to answer. This Action Bank restricts the Investigator to concrete, verifiable tools for gathering evidence, logging every action within the CaseState. The agent determines the necessary information, executes the corresponding action, and iteratively updates its hypotheses. Subsequently, a Challenger agent searches for alternative explanations and attempts to weaken the hypothesis, while a Method Critic scrutinizes the validity of the applied procedure. Finally, the Verifier cross-checks facts directly against the original records. Through this multi-faceted approach, the system goes beyond merely classifying transactions as fraudulent or non-fraudulent, instead constructing a fully traceable and defensible investigation detailing what was examined, what evidence was retrieved, and the rationale behind the final conclusion.

How we built it

The solution functions as an intelligent forensic auditing system that analyzes financial transactions and relationships to detect potential fraud without relying on a single classification model. Initially, data is stored in a SQLite database and queried through a repository that enables verifiable analysis. Based on these data, algorithms such as Bayesian networks and graph analysis generate observations and signals, which are subsequently converted into potential fraud hypotheses. An Investigator agent, supported by an LLM, analyzes the case status and determines what information needs to be verified, selecting specific actions from the Action Bank, such as querying transactions, contracts, vendors, employees, or bank movements. These actions are executed by Python directly against the database, ensuring that the LLM interprets and reasons without fabricating or calculating the evidence. The process is iterative: each new piece of evidence updates the CaseState, allowing the system to strengthen, weaken, or discard hypotheses. Subsequently, components such as the Challenger search for alternative explanations and the Method Critic reviews whether the reasoning contains errors or dependent evidence. Finally, a Verifier cross-checks the results directly against the original data, and an Evidence Gate determines whether sufficient, traceable evidence exists to issue a fraud finding; otherwise, the case may remain inconclusive. Thus, the architecture adheres to the principle that "the LLM interprets, Python calculates, and the database demonstrates," producing an auditable investigation rather than a simple binary "fraud" or "non-fraud" prediction.

Challenges we ran into

One of our primary challenges involved establishing a robust framework for data processing and analysis, defining precise criteria for fraud detection, and ensuring the consistency and reliability of agent evaluations. Furthermore, because this initiative marked our first implementation of machine learning within a predictive model, the team had to rapidly acquire both theoretical knowledge and practical expertise in this domain. Another significant obstacle was achieving team alignment. Since individual members held distinct visions regarding the architecture and development of the model, successfully harmonizing these diverse perspectives required deliberate coordination and consensus-building.

Accomplishments that we're proud of

A key achievement was clearly structuring the agent's architecture to integrate a machine learning process, enabling the AI to learn from its own errors. Despite our initial lack of prior experience in machine learning, we successfully conceived a model that analyzes patterns and data, learns from its mistakes, and delivers reliable judgment to the user.

What we learned

Our technical scope covered a diverse array of advanced domains, ranging from algorithms, machine learning models, individual agents, and multi-agent systems to graph theory. Additionally, we mastered full-stack development concepts, including frontend generation, the integration of frontend and backend architectures, and the training of neural networks.

What's next for LedgerLens

Future developments aim to enable natural conversational interactions between the user and the agent, leveraging advanced algorithms such as Isolation Forest for anomaly detection. Furthermore, the system is designed to dynamically adapt to provided variables and accurately interpret and translate them into actionable insights.