Daemon provides FinCrime Dynamics with an AI-driven solution that systematically gathers financial scam data from public sources such as Reddit, using AWS Bedrock's Foundation Models to augment FinCrime Dynamic's knowledge graph insights engine.

The Client

FinCrime Dynamics specialises in leveraging AI alongside financial crime and fraud intelligence data to help organisations build robust defences against illicit activities. They offer a suite of services aimed at identifying and closing critical security gaps across fraud and financial crime systems and processes.

The Challenge

Daemon first worked with FinCrime Dynamics as a part of a discovery/enablement process, sponsored by AWS. Our collaboration focused on understanding how AWS' advanced AI toolkit, including Amazon Bedrock, could be applied to FinCrime Dynamics specific advanced use cases, in addition to enabling and upgrading FinCrime Dynamics with advanced Generative AI prompt engineering and knowledge extraction techniques.

One critical use case uncovered was the need to generate new scenarios as part of the Scam Hunter product to keep the library of scam simulations up to date, relevant to FinCrime Dynamic's end-users, and validated against real data. This can be achieved by leveraging Generative AI to analyse financial crime case studies to generate fully-fledged simulations.

After the discovery period, FinCrime Dynamics and Daemon decided that the highest value idea to take to a POC (Proof of Concept) was a use case emerging from FinCrime Dynamics' most recent efforts in and around AI knowledge representation in fraud. The use case involved acquiring publicly accessible information on fraud and scams from diverse online platforms like LinkedIn, Facebook, and Reddit. The objective of this POC, also sponsored by AWS, is to systematically collect this vast and disparate data to construct a comprehensive knowledge graph and connect it to existing institutional knowledge resources. This graph, built on a specific financial crime ontology, is intended to capture the actors and timelines associated with scams and provide insights through sophisticated linking across public sources, institutional repositories, and customers' own data. Insights from this include rapidly identifying new scam trends, validating the models that FinCrime dynamics offer for generating synthetic scam data, and assisting in building out a robust library of fraudulent behaviours and their key characteristics. 

The POC process involved building an initial prototype that incorporated best practices and tooling to gain early insight into the approach's validity. This also set FinCrime Dynamics up to build upon and integrate it into their existing products. 

Our Approach

During the discovery/enablement phase, in addition to the discovery process, Daemon assisted FinCrime Dynamics in integrating Generative AI into their existing systems. In particular, best practices in prompt engineering and agentic workflows, such as few shot prompting, decomposition, and chaining, were demonstrated to assist FinCrime Dynamics in building out their proof of concept. 

The discovery process also included a series of short workshops, including impact/effort assessments, to identify the highest-value AI-driven use cases for FinCrime Dynamics. The process went beyond the technical and involved working together on the value propositions of different use cases with respect to FinCrime Dynamic's core business.

The second phase, centred on the POC of an insights-generating knowledge-graph from public sources. Our methodology consisted of an AI-enabled automated framework for the acquisition and synthesis of intelligence from unstructured public data sources, focusing on Reddit and LinkedIn in the first instance. This process was orchestrated by a Large Language Model (LLM, AWS Bedrock with Claude Sonnet 3.5 v2), which performed continuous semantic analysis to assess the relevance of discovered content against our target queries, yielding potential candidates for further analysis. 

The framework then performed a deep extraction of multi-modal data, capturing text, imagery, and metadata from nested and interactive elements to ensure contextual integrity. In the final synthesis stage, the aggregated, unstructured data was processed by an advanced LLM (AWS Bedrock with Claude Sonnet 3.5 v2). The model analysed the raw content to infer complex relationships between entities, structuring the information using FinCrime Dynamics' powerful graph ontology, to create an interconnected, machine-interpretable representation of scam information.

The Outcome

AI-enabled team: FinCrime Dynamics' data science team was upgraded in the emerging area of Generative AI, increasing their productivity and agility in developing new capabilities for both new and existing products.

Prioritised Generative AI use-cases: As a result of the discovery around use case selection, FinCrime Dynamics were able to improve their decision-making around product and technology. 

Evaluation for improved decision making: FinCrime Dynamics was able to conduct a preliminary evaluation of the proposed knowledge-graph generation tool.

Foundation for competitive advantage via AI expansion: The contributed solution is a starting point for a fully-fledged knowledge-graph generation tool whose linking of disparate and public data sources, carries a value proposition conferring competitive advantage, containing:

• Enhanced insights into client incidents
• Enhanced fraud pattern identification
• Validation of existing data repositories
• Improved synthetic data generation
  
  

Testimonials

Working with the Daemon team has been a genuine pleasure. Over a series of focused workshops, they quickly grasped the nuances of our domain, and worked with us to define a value-adding scope. From the outset, they demonstrated deep technical expertise, which enabled us to get everything operational far sooner than expected.

Daniel Turner-Szymkiewicz, Founder, FinCrime Dynamics