SF Machine Learning Meetup

Virtual. Register for the Zoom.

Foundational capabilities and models for generalist agents for computers

As we move toward a future where language agents can operate software, browse the web, and automate tasks across digital environments, a pressing challenge emerges: how do we build foundational models that can act as generalist agents for computers? In this talk, we explore the design of such agents—ones that combine vision, language, and action to understand complex interfaces and carry out user-intent accurately.

We present OmniACT as a case study, a benchmark that grounds this vision by pairing natural language prompts with UI screenshots and executable scripts for both desktop and web environments. Through OmniACT, we examine the performance of today’s top language and multimodal models, highlight the limitations in current agent behavior, and discuss research directions needed to close the gap toward truly capable, general-purpose digital agents.

About the Speaker

Raghav Kapoor is a machine learning at Adobe, where he works on the Brand Services team, contributing to cutting-edge projects in brand intelligence. His work blends research with machine learning, reflecting his deep expertise in both areas. Prior to joining Adobe, Raghav earned his Master’s degree from Carnegie Mellon University, where his research focused on multimodal machine learning and web-based agents. He also brings industry experience from his experience as a strategist at Goldman Sachs India.

BEARCUBS: Evaluating Web Agents' Real-World Information-Seeking Abilities

The talk focuses on the challenges of evaluating AI agents in dynamic web settings, the design and implementation of the BEARCUBS benchmark, and insights gained from human and agent performance comparisons. In the talk, we will discuss the significant performance gap between human users and current state-of-the-art agents, highlighting areas for future improvement in AI web navigation and information retrieval capabilities.

About the Speaker

Yixiao Song is a Ph.D. candidate in Computer Science at the University of Massachusetts Amherst. Her research focuses on enhancing the evaluation of natural language processing systems, particularly in assessing factuality and reliability in AI-generated content. Her work encompasses the development of tools and benchmarks such as VeriScore, an automatic metric for evaluating the factuality of long-form text generation, and BEARCUBS, a benchmark for assessing AI agents' ability to identify factual information from web content.

Visual Agents: What it takes to build an agent that can navigate GUIs like humans

We’ll examine conceptual frameworks, potential applications, and future directions of technologies that can “see” and “act” with increasing independence. The discussion will touch on both current limitations and promising horizons in this evolving field.

About the Speaker

Harpreet Sahota is a hacker-in-residence and machine learning engineer with a passion for deep learning and generative AI. He’s got a deep interest in RAG, Agents, and Multimodal AI.

Implementing a Practical Vision-Based Android AI Agent

In this talk I will share with you practical details of designing and implementing Android AI agents, using deki.

From theory we will move to practice and the usage of these agents in
industry/production.

For end users - remote usage of Android phones or for automation of standard tasks. Such as:

• "Write my friend 'some_name' in WhatsApp that I'll be 15 minutes late"
• "Open Twitter in the browser and write a post about 'something'"
• "Read my latest notifications and say if there are any important ones"
• "Write a linkedin post about 'something'"

And for professionals - to enable agentic testing, a new type of test that only became possible because of the popularization of LLMs and AI agents that use them as a reasoning core.

About the Speaker

Rasul Osmanbayli is a senior Android developer at Kapital Bank, Baku/Azerbaijan. It is the largest private bank in Azerbaijan. He created deki, an Image Description model that was used as a foundation for an Android AI agent that achieved high results on 2 different benchmarks: Android World and Android Control.

He previously worked in Istanbul/Türkiye for various companies as an
Android and Backend developer. He is also a MS at Istanbul Aydin University in Istanbul/Türkiye.

Register for the Zoom

Part 1: Navigating the GUI Agent Landscape

Understanding the Foundation Before Building

The GUI agent field is evolving rapidly, but success requires an understanding of what came before. In this opening session, we'll map the terrain of GUI agent research—from the early days of MiniWoB's simplified environments to today's complex, multimodal systems tackling real-world applications. You'll discover why standard vision models fail catastrophically on GUI tasks, explore the annotation bottlenecks that make GUI datasets so expensive to create, and understand the platform fragmentation that makes "click a button" mean twenty different things across datasets.

We'll dissect the most influential datasets (Mind2Web, AITW, Rico) and models that have shaped the field, examining their strengths, limitations, and the research gaps they reveal. By the end, you'll have a clear picture of where GUI agents excel, where they struggle, and, most importantly, where the opportunities lie for your own contributions.

About the Instructor

Harpreet Sahota is a hacker-in-residence and machine learning engineer with a passion for deep learning and generative AI. He’s got a deep interest in RAG, Agents, and Multimodal AI.

Virtual. Register for the Zoom

Quiero ser parte del mundo de AI, como lo logro?

En esta charla, compartiré mi trayectoria personal hacia el mundo de la inteligencia artificial (IA), comenzando con mi formación como ingeniero electrónico y mi doctorado en neuroinformática. Destacaré cómo mi tesis laureada sobre modelos volumétricos realistas para la localización precisa de fuentes EEG abrió puertas a oportunidades en procesamiento digital y visión 3D. Con experiencia docente en la Universidad Nacional de Colombia y certificaciones en machine learning y deep learning, discutiré cómo estos hitos me llevaron a desempeñarme como desarrollador de currículo para DeepLearning.AI, ofreciendo valiosas lecciones para quienes deseen seguir un camino similar.

Presentador

Ernesto Cuartas es un ingeniero electrónico y PhD en neuroinformática. Tesis PhD laureada “Forward volumetric modeling framework for realistic head models towards accurate EEG source localization”. Profesor asociado Universidad Nacional de Colombia. Experto en implementación y desarrollo de proyectos en procesamiento digital de señales, procesamiento digital de imágenes, visión 3D, computación gráfica, geometría computacional, fotogrametría e inteligencia artificial. Con certificaciones profesionales en machine learning, deep learning y data engineering. Actualmente trabajo como curriculum developer/engineer para DeepLearning.AI.

Domina tus Datos Médicos: De la Curación al Impacto Clínico

Los datos de alta calidad son la base de un aprendizaje automático efectivo en el ámbito de la salud. Esta charla presenta estrategias prácticas y técnicas emergentes para gestionar datasets de imágenes médicas, desde la generación de datos sintéticos y la curación, hasta la evaluación y el despliegue.

Comenzaremos con casos de estudio reales de investigadores y profesionales que están transformando sus flujos de trabajo en imágenes médicas mediante prácticas centradas en los datos. Luego pasaremos a un tutorial práctico utilizando FiftyOne, la plataforma open-source para la inspección visual de datasets y la evaluación de modelos. Los asistentes aprenderán a cargar, visualizar, curar y evaluar datasets médicos en distintos tipos de imágenes.

Ya seas investigador, clínico o ingeniero de ML, esta charla te brindará herramientas e ideas prácticas para mejorar la calidad de tus datos, la fiabilidad de tus modelos y su impacto clínico.

Presentadora

Paula Ramos tiene un doctorado en Visión Artificial y Aprendizaje Automático, con más de 20 años de experiencia en el campo tecnológico. Desde principios de la década del 2000 en Colombia, ha desarrollado novedosas tecnologías integradas de ingeniería, principalmente en Visión Artificial, robótica y Aprendizaje Automático aplicados a la agricultura.

Agentes AI Multi-Fuente y Embebidos

Demostraré cómo construir agentes de IA contextualmente conscientes, capaz de responder y tomar acciones entre multiples sistemas privados y la implementación de RAG semántico a través de fuentes de datos dispares, embebidos en sistemas existentes, todo esto sin necesidad de una infraestructura compleja de MLOps.

Presentador

Kevin Blanco es un Senior DevRel Advocate, Charlista Internacional con más de 15 años en liderazgo tecnológico. Ha diseñado estrategias de IA en IBM Watson y desarrollado soluciones para Google, Microsoft y Nintendo.

Más allá del modelo: Metodología y buenas prácticas para liderar proyectos exitosos de IA con CPMAI

El éxito de los proyectos de IA no depende solo del modelo o de los datos, sino de cómo se gestionan desde el inicio. En esta charla exploraremos la metodología CPMAI (Cognitive Project Management for AI) avalada por el Project Management Institute - PMI, un marco estructurado que permite a los equipos de IA alinear sus iniciativas con objetivos de negocio, gestionar riesgos éticos y mejorar los resultados. Compartiremos buenas prácticas que pueden ser adaptadas por profesionales técnicos para mejorar la entrega de valor en cada fase del proyecto e implementar soluciones de IA éticas y responsables.

Presentadora

Ivonne Mejía B. es especialista en gestión de proyectos tecnológicos, con más de 20 años de experiencia internacional en el sector privado y académico en México, Canadá y Estados Unidos. Está certificada en CPMAI™, PMP®, Prosci®, y cuenta con un diplomado en Liderazgo Tecnológico por UC Berkeley. Disfruta colaborar, aprender en comunidad y compartir su experiencia para ayudar a las organizaciones a definir estrategias de transformación con IA y liderar soluciones éticas y responsables.

Register for the Zoom

Part 2: From Pixels to Predictions - Building Your GUI Dataset

Hands-On Dataset Creation and Curation with FiftyOne

The best GUI models are only as good as their training data, and the best datasets are built by understanding what makes GUI interactions fundamentally different from natural images. In this practical session, you'll build a complete GUI dataset from scratch, learning to capture the precise annotations that GUI agents need.

Using FiftyOne as your data management backbone, you'll import diverse GUI screenshots, explore annotation strategies that go beyond bounding boxes, and implement efficient labeling workflows. We'll tackle the real challenges: handling platform differences, managing annotation quality, and creating datasets that transfer to new domains. You'll also learn advanced techniques like synthetic data generation and automated prelabeling to scale your annotation efforts.

Walk away with a production-ready dataset and the skills to build more—because in GUI agents, data quality determines everything.

By the end, you'll have both a dataset and the methodology to build the next generation of GUI training data.

About the Instructor

Harpreet Sahota is a hacker-in-residence and machine learning engineer with a passion for deep learning and generative AI. He’s got a deep interest in RAG, Agents, and Multimodal AI.