Current Research Projects

Advancing SLMs

Aligning LM Pretraining Data with Downstream Uses

This project focuses on advancing small language models (SLMs), a rapidly emerging paradigm that is shaping the future of AI and enabling agentic systems designed for task-specific applications. Unlike traditional large-scale models, SLMs promise efficiency, adaptability, and scalability, making them ideal for research into real-world deployments in financial domains. The project aims to curate and reformat pretraining datasets so that smaller models can be studied for their ability to not only outperform much larger counterparts but also substantially reduce hallucinations, improving reliability and trustworthiness in autonomous AI systems.

Causal AI

Cause-Effect Inference for Task-Oriented AI Training

This project aims to explore drivers of causal inference in unstructured text. Causal inference is a cornerstone of evidence-based decision-making in complex environments. This initiative aims to identify and quantify the impact of specific phenomena derived from natural language on outcomes, enabling multiple financial applications, such as understanding the impact of different words and sentences from public corporate communications (e.g., earning calls) and observed market behaviours in retrospective analyses.

Responsible AI

Safeguarding Compliant Responses

This initiative advances responsible AI by developing reliable, context‑aware safeguards for conversational AI and chatbot systems used in high‑stakes, real‑world settings. As large and small language models are increasingly embedded in workflows, supporting efforts toward consistent, safe, and trustworthy behavior across diverse contexts is essential. By transforming policies and norms into actionable constraints, this initiative investigates approaches to policy‑aware reasoning, explores methods to reduce hallucinations, and supports scalable, trustworthy conversational AI in research and prototype environments.

Cognitive AI

Emotion-Aware Conversational Adaptation

This research integrates emotional intelligence into conversational AI, with the goal of enabling interactions that are not only accurate but also empathetic, adaptive, and trustworthy. In high‑stakes conversations where users may experience stress or uncertainty, emotion‑aware response adaptation is critical for effective support. The work focuses on detecting contextual and linguistic indicators of emotional state and dynamically adjusting model behavior to ensure clarity, empathy, and alignment with user preferences—without inferring, diagnosing, or labeling individual mental or emotional conditions.