Hierarchical Reason-of-Contact Detection in Retail Banking Customer Interactions via LLM-Driven Taxonomy Induction
ACL 2026 Industry Track (Oral)
Yoages Kumar Mantri, Yagneswaran Sriraja, John Hartley, Greig A. Cowan, Raad Khraishi, Abhinav Gupta
Introduces a deployed, hierarchical Reason-of-Contact detection system for retail banking conversations. The approach combines clustering, LLM-driven taxonomy induction, and rapid retrieval to identify customer intent across channels, supporting real-time performance, auditability, dynamic taxonomy updates, and improved detection of out-of-distribution contact reasons.
ㅤ
Helping Customers in Distress: An LLM-powered Agent that Converses, Probes, and Routes
ICLR 2026 Workshop: Advances in Financial AI
Alankar Atreya, Stefan Sylvius Wagner, Devesh Batra, Robert Hankache, Cristovão Iglesias Jr, Patrick Sinclair, Giulio Pelosio, Michael McMillan, Greig A. Cowan, Raad Khraishi
Presents a customer-facing LLM triage agent for fraud, scam, and dispute journeys. The agent conducts multi-turn conversations, asks targeted follow-up questions, and routes cases in line with policy, with evaluation using historical cases and synthetic customer digital twins.
ㅤ
Large-Scale Chatbot Validation Through Customer Digital Twin Simulations
ICLR 2026 Workshop: Advances in Financial AI
Cristovão Iglesias Jr, Devesh Batra, Alankar Atreya, Stefan Sylvius Wagner, Robert Hankache, Patrick Sinclair, Giulio Pelosio, Michael McMillan, Greig A. Cowan, Raad Khraishi
Introduces a scalable validation framework for customer-facing chatbots using synthetic customer agents grounded in transactional and conversational data. The framework enables large-scale testing through automated evaluation, expert review, and adversarial probing before deployment.
ㅤ
AutoQA: An Interpretable Automation Framework for CDD Quality Assurance in Financial Services
ICLR 2026 Workshop: Advances in Financial AI
Iman Zafar, Andrea Bassani, Alexandros Anatolakis, Greig A. Cowan, Raad Khraishi
Presents an interpretable automation framework for customer due diligence quality assurance. AutoQA combines deterministic rules with targeted LLM-assisted reasoning, keeping analysts in the loop while improving consistency, interpretability, and review efficiency in regulated financial-services workflows.
ㅤ
Evaluating Performance Drift from Model Switching in Multi-Turn LLM Systems
ICLR 2026 Workshop: Catch, Adapt, and Operate: Monitoring ML Models Under Drift
Raad Khraishi, Iman Zafar, Katie Myles, Greig A. Cowan
Introduces a switch-matrix benchmark for measuring performance drift when LLM systems switch models mid-conversation. The paper shows that handoffs between models can affect reliability in multi-turn settings, motivating explicit monitoring and evaluation of model switching in production LLM systems.
ㅤ
MOSAIC: Modular Orchestration for Structured Agentic Intelligence and Composition
arXiv preprint
Yifan Bao, Xinyu Xi, Xinyu Liu, Wen Ge, Lei Jiang, Kevin Zhang, Raad Khraishi, Yihao Ang, Anthony K. H. Tung, Lukasz Szpruch, Hao Ni
A memory-grounded agentic framework that treats automated data science as structured model selection. MOSAIC retrieves prior cases and code modules to build a reusable blueprint specifying the chosen modelling components and how they compose, then validates candidates by execution and refines them with feedback. On financial time-series forecasting and generation, it improves task performance, execution success, and decision traceability over AutoML and agentic baselines.
Operationalising LLMs for Compliance-Critical Letter Writing in Financial Services
NeurIPS Workshop on Generative AI in Finance
Devesh Batra, Alexandros Anatolakis, John Hartley, Jude King, Greig Cowan, Raad Khraishi
Describes a deployed LLM system for drafting complaint resolution letters, including an end-to-end architecture and continuous compliance monitoring in a high-stakes setting.
ㅤ
A Review of LLM Agent Applications in Finance and Banking
SSRN preprint
Devesh Batra, Conor Hamill, John Hartley, Ramin Okhrati, Dale Seddon, Harvey Miller, Raad Khraishi, Greig Cowan
A survey of LLM agents in finance and banking, organizing applications into simulation, acting, analysis, and advising, and discussing technical, ethical, regulatory, and operational considerations.
ㅤ
Documenting Deployment with Fabric: A Repository of Real-World AI Governance
AAAI/ACM Conference on AI, Ethics, and Society (AIES)
Mackenzie Jorgensen, Kendall Brogle, Katherine M. Collins, Lujain Ibrahim, Arina Shah, Petra Ivanovic, Noah Broestl, Gabriel Piles, Paul Dongha, Hatim Abdulhussein, Adrian Weller, Jillian Powers, Umang Bhatt
Introduces Fabric, a public repository of deployed AI use cases and their governance mechanisms, built from practitioner interviews and co-designed workflow diagrams to study real-world oversight patterns.
ㅤ
Obscured but Not Erased: Evaluating Nationality Bias in LLMs via Name-Based Bias Benchmarks
arXiv preprint
Giulio Pelosio, Devesh Batra, Noémie Bovey, Robert Hankache, Cristovao Iglesias, Greig Cowan, Raad Khraishi
Proposes a name-based benchmarking approach (derived from BBQ) to measure nationality bias when explicit demographic labels are removed, reflecting real world settings where bias can remain latent.
ㅤ
Evaluating the Sensitivity of LLMs to Prior Context
arXiv preprint
Robert Hankache, Kingsley Nketia Acheampong, Liang Song, Marek Brynda, Raad Khraishi, Greig Cowan
Introduces benchmarks that systematically vary prior context in multi-turn settings and shows that performance can degrade substantially as context grows — while design choices like instruction placement can help mitigate drops.
ㅤ
AUTOSUMM: A Comprehensive Framework for LLM-Based Conversation Summarization
ACL (Industry Track)
Abhinav Gupta, Devendra Singh, Greig Cowan, N Kadhiresan, Siddharth Srivastava, Yagneswaran Sriraja, Yoages Kumar Mantri
Presents an LLM summarization system for customer-advisor conversations with privacy/compliance safeguards, including segmentation, coverage tracking, and multi-layer hallucination detection informed by human-in-the-loop feedback.
ㅤ
How Personality Traits Shape LLM Risk Taking Behaviour
Findings of the Association for Computational Linguistics (ACL)
John Hartley, Conor Brian Hamill, Dale Seddon, Devesh Batra, Ramin Okhrati, Raad Khraishi
Studies how LLM “personality” relates to risk-taking via Cumulative Prospect Theory, and explores how prompting-based trait interventions can shift model risk propensity.
ㅤ
Representation Learning on Large Non-Bipartite Transaction Networks using GraphSAGE
Graph-Based Representations in Pattern Recognition (GbRPR)
Mihir Tare, Clemens Rattasits, Yiming Wu, Euan Wielewski
Applies GraphSAGE to large banking transaction networks to learn embeddings that reveal interpretable structure and improve downstream prioritization in a money mule detection setting.
ㅤ
Agent-based Modelling of Credit Card Promotions
International Journal of Bank Marketing
Conor B. Hamill, Raad Khraishi, Simona Gherghel, Jerrard Lawrence, Salvatore Mercuri, Ramin Okhrati, Greig Cowan
Develops an agent based model of the UK credit card market to explore outcomes of promotion strategies under different competitive scenarios and market conditions.
A Brief Review of Quantum Machine Learning for Financial Services
arXiv preprint
Mina Doosti, Petros Wallden, Conor Brian Hamill, Robert Hankache, Oliver Thomson Brown, Chris Heunen
Reviews quantum machine learning methods with potential applications in finance, and discusses practical challenges, opportunities, and limitations.
Conformal Predictions for Longitudinal Data
arXiv preprint
Devesh Batra, Salvatore Mercuri, Raad Khraishi
Introduces Longitudinal Predictive Conformal Inference (LPCI), a distribution-free method for prediction intervals in longitudinal datasets that targets both cross-sectional and longitudinal coverage.
ㅤ
Modelling customer lifetime value in the retail banking industry
arXiv preprint
Greig Cowan, Salvatore Mercuri, Raad Khraishi.
Presents a machine learning framework for customer lifetime value forecasting over arbitrary horizons in retail banking, with a production implementation and demonstrated improvements over baseline approaches.
Offline Deep Reinforcement Learning for Dynamic Pricing of Consumer Credit
ACM International Conference on AI in Finance (ICAIF)
Raad Khraishi, Ramin Okhrati
Uses offline deep reinforcement learning to learn personalized credit pricing policies from static datasets, avoiding risky online experimentation while enabling policy improvement.
ㅤ
An Introduction to Machine Unlearning
arXiv preprint
Salvatore Mercuri, Raad Khraishi, Ramin Okhrati, Devesh Batra, Conor Hamill, Taha Ghasempour, Andrew Nowlan
A survey of machine unlearning: consolidates definitions, compares key algorithms, and discusses evaluation and implementation challenges for real deployments.
<hr><p>Discover our AI research was originally published in NatWest Group AI & Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>