AI has the potential to transform learning, work, and daily life for millions of people, but only if we design with accessibility at the core. Too often, disabled people are underrepresented in datasets, creating systemic barriers that ripple through models and applications. This talk explores how data scientists and technologists can mitigate bias, from building synthetic datasets to fine-tuning LLMs on accessibility-focused corpora. We’ll look at opportunities in multimodal AI: voice, gesture, AR/VR, and even brain-computer interfaces, that open new pathways for inclusion. Beyond accuracy, we’ll discuss evaluation metrics that measure usability, comprehension, and inclusion, and why testing with humans is essential to closing the gap between model performance and lived experience. Attendees will leave with three tangible ways to integrate accessibility into their own work through datasets, open-source tools, and collaborations. Accessibility is not just an ethical mandate, it’s a driver of innovation, and it begins with thoughtful, human-centered data science.

Large language models are powerful, but their true potential emerges when they evolve into AI agents which are systems that can reason, plan, and take action autonomously. My talk will explore the shift from using models as passive tools to designing agents that actively interact with data, systems, and people.

I will cover: - Gen AI and Agentic AI – How are These Different - Single Agent (monolithic) and Multi Agent Architectures (modular / distributed) - Open Source and Closed Source AI Systems - Challenges of Integrating Agents with Existing Systems

I will break down the technical building blocks of AI agents, including memory, planning loops, tool integration, and feedback mechanisms. Examples will be used to highlight how agents are being used in workflow automation, knowledge management, and decision support.

I will wrap up with where limitations of AI Agents still pose risks: - Assessing Maturity Cycle of Agents - Cybersecurity Risks of Agents

By the end, attendees will understand: - What makes AI agents different from LLMs - Technical considerations required to build AI Agents responsibly - Applicable knowledge to begin experimenting with agents.

A central challenge in knowledge transfer lies in the transfer of tacit knowledge. LLMs, capable of identifying latent patterns in data, present an interesting opportunity to address this issue. This paper explores the potential of LLMs to externalize experts’ tacit knowledge and aid its transfer to novices. Specifically, we examine three questions: RQ1: Can LLMs effectively externalize experts’ tacit knowledge? How to do so (e.g., prompting strategy)? RQ2: How can LLMs use externalized tacit knowledge to make effective decisions? RQ3: How can LLM-externalized tacit knowledge support novice learning? We explore these questions using real-world tutoring conversations collected by Wang et al. (2024).

Our findings suggest that LLMs may be capturing nuances from experts’ observed behavior that are different from the knowledge experts articulate. With carefully designed prompting strategies, LLMs may offer a practical and scalable means of externalizing and transferring tacit knowledge.