Focus Areas

My work is organized around a simple question: how can AI systems help people make better decisions in complex real-world settings? The areas below describe the product ideas, research directions, and contributions behind that question.

Agentic AI Products

Knowledge work rarely happens in a single prompt. It unfolds across documents, tools, decisions, interruptions, and follow-up actions. We build agentic products that help people keep this work moving without losing context.

At AI4Alpha / UCoWorker, our team turns the agent idea into a usable product: shaping how agents remember work, resume tasks, coordinate tools, and present results across product surfaces. My contribution focuses on product workflow design, agent behavior, and implementation across the product stack.

Try our products: AI4Alpha and UCoWorker.

Decision-Making for Finance

Financial-service operations involve repeated decisions: when to act, when to wait, what message to send, and how to balance value, cost, timing, and customer experience. Our team works on AI systems that treat these choices as decision problems rather than isolated predictions.

Within the team, I helped connect historical interaction data with practical decision workflows. My work covered data preparation, model evaluation, deployment support, experiment analysis, and reliability checks, with the goal of making learned policies useful and accountable in operational settings.

This direction connects reinforcement learning, experimentation, and causal reasoning, but the public-facing question is broader: does the system recommend actions that make sense when real costs and constraints matter?

Spatial-Temporal Representation Learning

Many real systems, such as traffic, buildings, energy, and infrastructure, change over both space and time. My PhD research studies how models can learn from long histories of system behavior instead of relying only on recent observations.

My contribution is developing representation-learning methods that make historical patterns easier to reuse for forecasting and decision support. This includes work on long-term multivariate history representation and retrieval over past system trajectories.

Representative publications: TAI 2026 and TNNLS 2025.

Graph Learning for Dynamic Systems

Real-world systems are connected: roads influence nearby roads, buildings interact with weather and occupancy, and sensors often describe only part of a larger system. Graph learning gives these relationships a structure that models can reason over.

My work in this area focuses on learning useful relationships from data when the underlying system is dynamic and multi-scale. The goal is to make models more aware of how local changes propagate through larger systems.

Representative publication: IS 2024.

Applied Machine Learning for Energy and Control

Forecasting is most useful when it supports action. In energy and building systems, this means connecting prediction with control, safety, and operational constraints.

My applied ML work studies how learning methods can support building energy management, safe coordination, and adaptive control. I contribute by connecting algorithm design with evaluation settings that reflect how these systems are actually used.

Representative publications: IoTJ 2026, EPEC 2025, and IEEE Access 2022.

LLM Evaluation

As language models become more capable, it becomes more important to understand how they behave across languages, prompts, and use cases. Our team is interested in evaluation work that makes model behavior easier to inspect, compare, and improve.

Our current study evaluates implicit and explicit gender stereotypes in LLM families. I contribute to the evaluation design, model comparison, statistical analysis, and reproducible workflow. The manuscript is under review.