AI in Financial Services: From LLM Failure Modes to Graph-Based Fraud Detection

The second session of Tokyo AI's Financial Services series digs into the engineering realities of running AI inside high-stakes financial environments. Three talks cover where large language models fail in quantitative and trading workflows, how vertical multi-model AI systems can structure messy unstructured institutional investment data, and how Knowledge Graphs combined with Graph Neural Networks support network-level fraud and anomaly detection. Colin Rowat (Rakuten Institute of Technology) opens with the ways AI and LLMs go wrong in finance, drawing on examples from hedge funds, trading systems, and benchmarks, and why leading firms stay cautious about production deployment. Jeff Tsui (Visual Alpha) walks through the reconciliation problem across back-office, middle-office, and front-office data, and what a multi-model architecture takes to reach institutional-grade extraction accuracy. Sanjeev Sinha (SHIFT) closes with an agentic architecture that uses graph topology, GraphSAGE and GATv2, and vector similarity search to surface fraud that is invisible in isolation. The session is aimed at researchers, engineers, and technical leaders who want concrete detail on building production-grade financial AI infrastructure. Doors open at 18:00, talks run from 18:30, and networking follows from 20:00.