Titulo Estágio
Development of an AI-Powered Insights Agent for Fraud Case Management
Local do Estágio
Remote, Hybrid, or Lisbon/Coimbra/Porto offices.
Enquadramento
Fraud analysts routinely interact with our internal "Case Manager" tool, which holds a wealth of data about fraud cases. Beyond individual case details, there's a critical need for actionable insights derived from this data to inform the risk strategy, evaluate performance, and drive analytical improvements. Manually extracting and synthesizing this higher-level strategic information is time-consuming and often requires specialized analytical skills. An AI Agent capable of autonomously generating insights related to the risk strategy, performance metrics, and broader analytics within the Case Manager would significantly empower fraud analysts and risk managers. This project aims to develop such an "Insights Agent" leveraging advanced AI and Large Language Models (LLMs) to provide real-time, context-aware intelligence, ultimately enhancing proactive fraud prevention and strategic decision-making.
Objetivo
● To design and develop an AI Agent, "Insights Agent", capable of extracting and synthesizing strategic information from the "Case Manager" tool related to fraud risk strategy, performance, and analytics.
● To implement advanced AI models, including Large Language Models, to identify patterns, anomalies, and trends within fraud case data, providing actionable insights.
● To evaluate the effectiveness of the "Insights Agent" in terms of providing relevant and accurate strategic insights, reducing manual analysis time for analysts, and positively impacting risk strategy and operational performance.
● To explore different AI and NLP techniques, including LLM fine-tuning and
benchmarking, for optimal insight generation and hallucination prevention.
● To analyze hallucination prevention, including theoretical basis, implementation, and
impact on insight accuracy. Investigate fact verification and robust evaluation.
Analyze trade-offs (cost, complexity, effectiveness) and integration into pipelines,
considering combined techniques for better solutions.
Plano de Trabalhos - Semestre 1
Month 1-2: Requirements Gathering and Data Understanding
○ Deep dive into the "Case Manager" tool to understand its data architecture,
available data points, and the current methods fraud analysts and risk
managers use to derive insights.
○ Collaborate with risk consultants and fraud analysts to define key strategic
insights needed (e.g., performance trends, risk hot spots, strategy
effectiveness metrics).
○ Identify and define data sources within Case Manager relevant for generating
risk strategy, performance, and analytical insights.
● Month 3-4: Literature Review and Initial Model Development
○ Extensive research on state-of-the-art AI techniques for insight generation,
data interpretation, and multi-modal data analysis, with a focus on leveraging
LLMs for complex data synthesis.
○ Set up a development environment and necessary infrastructure for data
ingestion, processing, and AI model training.
○ Implement initial data pipelines to extract and structure relevant data from
Case Manager for analytical processing.
● Month 5-6: Prototype Development and Baseline Evaluation
○ Develop a basic prototype of the "Insights Agent" focusing on a specific set of
critical insights (e.g., top fraud typologies, performance metrics for a given
period).
○ Conduct preliminary evaluations of the generated insights for accuracy,
relevance, and utility with input from fraud analysts and risk managers.
Plano de Trabalhos - Semestre 2
● Month 7-8: Advanced Insight Generation
○ Enhance the "Insights Agent" to generate more complex and nuanced
strategic insights (e.g., impact of strategy changes, predictive analytics for
emerging risks, root cause analysis of performance drops).
● Month 9-10: Model Refinement
○ Explore advanced techniques for causality detection, anomaly detection, and
natural language query processing over structured and unstructured data.
○ Refine AI models, including LLM prompts and architectures, to improve the
quality, accuracy, and depth of the insights, with a strong focus on
hallucination prevention and fact verification.
● Month 11-12: Comprehensive Testing and Documentation
○ Conduct extensive testing of the "Insights Agent," including functional,
performance, security, and usability testing.
○ Develop detailed documentation for the "Insights Agent," covering technical
specifications, user manuals, maintenance procedures, and a guide for
interpreting the generated insights.
Condições
● Access to "Case Manager" tool and relevant data for development and testing, with
appropriate data privacy and security measures in place.
● Provision of a development environment with necessary computational resources.
● Regular mentorship and guidance from the internal team, including fraud, risk, and
technical experts.
● Opportunity to publish research findings, respecting company confidentiality
agreements.
Observações
This project represents a significant leap towards
augmenting human intelligence with AI in the critical domain of fraud management. By
providing proactive and data-driven insights, the "Insights Agent" will not only improve the
efficiency of fraud analysis but also enable more strategic and agile responses to evolving
fraud threats. This initiative aligns perfectly with Feedzai's commitment to leveraging
cutting-edge AI to combat financial crime and could become a core component of our future
analytical capabilities.
Orientador
Javier Liébana de la Barrera
javier.liebana@feedzai.com 📩