BayNode - The Bay Area Node.js Meetup

Oct 22, 2026
9:00 AM - 11:00 AM PST
Online. Register for the Zoom!

Multi-Agent Communication: A framework, diagnostic and mechanistic perspective

Multi-agent LLM systems are increasingly used for collaborative reasoning, debate, and consensus, yet their communication dynamics remain poorly understood. This talk presents a framework for studying multi-agent communication through diagnostic and mechanistic perspectives.

I will discuss CONSENSAGENT, which improves consensus by mitigating sycophancy, alongside our diagnostic work on communication patterns and failure modes in real-world multi-agent debates. I will then present ongoing work that moves toward a mechanistic understanding of how these interaction patterns arise internally, with the broader goal of making multi-agent systems more interpretable, reliable, and controllable.

About the Speaker

Priya Pitre I am an Ph.D student in the Computer Science Department at Virginia Tech (VT), co-advised by Dr. Xuan Wang and Dr. Naren Ramakrishnan.

Exposing and Improving Fine-Grained Visual Grounding Abilities of Lightweight Multimodal LLMs

Lightweight multimodal LLMs can localize whole objects effectively, yet often struggle when a query targets a small object part or fine-grained visual detail. This talk presents a reasoning-guided framework that teaches compact models to ground parts through an explicit coarse-to-fine process: first locating the parent object, then identifying the requested part.

A part-aware reinforcement-learning objective provides stage-wise rewards for object accuracy, part containment, and the consistency of the model’s self-critique. Using these techniques, a compact 4B-parameter model achieves state-of-the-art zero-shot part grounding while preserving its object-level performance.

These advances can be used to enable lightweight MLLMs to support detail-oriented tasks in biology and robotics.

About the Speaker

Kazi Mehrab I a CS PhD candidate at Virginia Tech, where I currently focus on multimodal LLMs and computer vision tasks, including visual perception, reasoning and grounding.

Understanding Visual Generative Models for Precise Control

Despite remarkable progress in image and video generation, translating user intent into precise and consistent visual outputs remains a challenge. This talk explores how understanding the representations within generative models can enable finer control over what they create.

It connects semantic image editing with compositional generation, examining how visual concepts can be isolated, manipulated, and combined while preserving their identity and surrounding content. Building on these insights, structured visual inputs provide a way to express complex intent through subject references, poses, and spatial layouts.

The discussion then extends from images to video, where representations must evolve to preserve scene continuity while accommodating motion and change. Together, these directions establish a unified perspective on how visual representations can support controllable editing, composition, and coherent generation across space and time.

About the Speaker

Yusuf Dalva is a Ph.D. candidate at Virginia Tech, advised by Pinar Yanardag and affiliated with the Sanghani Center for Artificial Intelligence and Data Analytics.

Oct 29, 2026
9:00 AM - 10:00 AM PST
Online. Register for the Zoom!

We’ll walk through how modern agentic systems move beyond simple prompts—leveraging structured tools like dataset operations, embeddings, evaluation pipelines, and model execution to take real action. You’ll see how these agents can tag data, run inference, evaluate performance, and surface insights automatically, all within a unified workflow.

By combining natural language interfaces with programmable building blocks, teams can dramatically reduce manual effort, accelerate experimentation, and unlock faster decision-making across the ML lifecycle.

Whether you're building data-centric AI systems, managing large-scale vision datasets, or exploring agentic workflows for the first time, this session will give you a practical blueprint for getting started.

Nov 05, 2026
9:00 AM - 11:00 AM PST
Online. Register for the Zoom!

Talks will include:

U-Net Framework for Micro-Scale Surface Damage Segmentation in High-Resolution Soybean Seed Imagery

Accurate detection of surface-level seed damage is critical for soybean seed quality assurance, yet automated micro-scale damage detection in high-resolution imagery remains an open challenge due to extreme spatial variability in damage scale, severe class imbalance across damage types, and the computational demands of processing ultra-high resolution agricultural imagery at scale. This research addresses the semantic segmentation of fine-grained soybean seed surface defects, such as wrinkles, dark spots, and general surface damage, in 6048 × 4024 pixel images where target damages can be as small as 18 × 18 pixels, using an augmented dataset of 77,000 individual seed images.

To overcome the difficulties of micro-scale detection, spatial sparsity, and class confusion, we propose a two-stage training and dual model inference framework built on an optimized U-Net architecture with a ResNet34 encoder. In the first stage, a damage specialist model is trained using weighted loss functions and a class-balanced approach that prioritizes damage classes.

In the second stage, transfer learning is applied to initialize a healthy seed specialist model from Stage 1 weights, with rebalanced class weights and tile probabilities that identify healthy seeds. During inference, a confidence-gated damage filter suppresses low-confidence predictions, and healthy seed labels are assigned only when the specialist model's confidence exceeds that of the damage model.

The two specialist models achieve validation accuracies of 94% and 98.53%, respectively, and the combined inference system successfully detects and localizes all three damage categories across unseen test images under conditions of extreme spatial variability and class imbalance. Qualitative evaluation confirms close alignment of predicted boundaries with ground truth annotations across all damage types, including dark spots.

These results demonstrate that confidence-gated dual model inference can reduce class confusion in imbalanced micro-scale segmentation, advancing the feasibility of fine-grained automated seed quality inspection at the scale.

About the Speaker

Saurav Upadhyaya is an AI/ML researcher with a Master's degree in Computer Science, whose work spans agricultural AI, conversational systems, and public health, consistently translating advanced technology into tangible real-world impact.

From camera to drone: wildfire detection and integration at the edge

Wildfires are getting larger and more expensive every year. But caught early, a fire is a small job for a small team; caught late, nothing stops it. The cameras to catch them are already installed, so the work is not necessary more sensors but making the ones we have act.

This talk walks the whole chain: a small vision model running on the camera that flags smoke, an integration layer that scores the alert against terrain, weather and history to decide whether it matters, and a drone dispatched to confirm before anyone commits a crew. We will cover what has to run at the edge and why, how small the model can get before it stops seeing smoke, and what it takes to trust a detection nobody has looked at yet.

About the Speaker

Maxime Carriere is co-founder of Kernwerk in Berlin, where he works on compressing AI models to run on cheap embedded hardware.

5,000 Flights to Answer One Question: Can a Drone Sample as Well as a Human?

Environmental testing still begins with a person walking onto a site with a shovel. We built a system that turns a plain-language brief into a sampling campaign, flies it, and returns soil, water and vegetation samples with GPS, timestamps and chain of custody intact: 5,000 flights, roughly 9,000 miles and 570 flight hours over a 185-acre site in Roatan, much of it under canopy where GPS degrades and clearances are tight.

I will walk through the vision and planning stack, including multispectral and thermal canopy sensing, terrain and land-cover mapping, and the constraint solver that clears a route against airspace and battery budget before anything leaves the ground. Then the harder problem: evidentiary comparability.

A drone-collected result means nothing to a regulator unless it agrees with a hand-collected split under the same method, so I will show the dual-collection protocol we designed to test exactly that. I will close on why calibrated uncertainty and knowing when a model should refuse to answer matter more than raw accuracy once the output lands in a regulatory filing.

About the Speaker

Rohan Talwadia is co-founder and CTO of LabGPT, which automates environmental field sampling using drones and edge analysis, with results returned through ISO/IEC 17025-accredited partner laboratories.

Nov 11, 2026
9:00 AM - 11:00 AM PST
Online. Register for the Zoom!

Agentic RAG: Beyond Retrieve-and-Generate

Retrieval-Augmented Generation (RAG) has become the default pattern for grounding large language models in external knowledge, but most implementations still follow a rigid retrieve-once, generate-once pipeline — one that struggles with multi-hop questions, ambiguous queries, and knowing when its own retrieved context is insufficient. This talk introduces Agentic RAG, where retrieval is treated as an action within an agent's reasoning loop rather than a fixed upstream step.

We'll examine query decomposition and routing for breaking complex questions into targeted sub-queries, self-reflective and corrective retrieval loops that let an agent judge and re-query its own results, and tool-orchestration patterns (via MCP) that let retrieval sit alongside other agent actions like database lookups and API calls. Using a live architecture — evolving a standard RAG chatbot into an agentic, MCP-connected system — we'll walk through what changes in design, and where these systems introduce new failure modes: grounding drift, latency and cost from repeated retrieval loops, and cases where a simpler RAG pipeline still wins.

Attendees will leave with a practical framework for deciding when the added complexity of agentic RAG is worth it, and a set of design patterns for building it correctly.

About the Speaker

Balaji Venkatasubramaniyar is a Technical Lead at Wisdom Infotech, leading a 15+ person engineering team delivering enterprise solutions. With 13+ years of experience in enterprise software and insurance technology, he specializes in agentic AI systems, RAG architectures, and vector databases.

GeoAI for the Physical World: Earth Observation, Foundation Models, and Urban Digital Twins

Earth observation provides a unique form of computer vision for understanding the physical world at city to continental scales. In this talk, I will show how satellite imagery, geospatial data, machine learning, and foundation-model representations can be combined to characterize urban environments and environmental conditions.

I will present UrbanScope Open, an open GeoAI digital-twin prototype integrating Earth observation with 3D buildings, vegetation, land-surface temperature, air quality, noise, population, and other urban data. I will also share lessons from my research using geospatial foundation-model embeddings for environmental prediction across Europe.

The talk will discuss how these approaches can contribute to increasingly multimodal AI systems capable of reasoning about real-world environments.

About the Speaker

Cesar Alvarez is a researcher at the University of Augsburg working at the intersection of GeoAI, Earth observation, remote sensing, and environmental intelligence. His research applies machine learning, computer vision, and geospatial foundation models to problems including urban environments, climate risk, air quality, and agriculture.

Can agents get curious?

Most AI agents are good at answering a question once we tell them exactly what to look for. The harder problem is building agents that can explore a complex dataset autonomously: generating hypotheses, deciding which analyses are worth running, allocating additional compute when evidence is ambiguous, and knowing when they have enough evidence to stop.

In this talk, I’ll show an architecture for autonomous research agents that combines structured knowledge, iterative tool use, and explicit evidence tracking to turn open-ended questions into a sequence of testable investigations. I’ll discuss practical lessons from building and evaluating these systems, including why more test-time compute does not automatically produce better research and how provenance and evaluation can make long-running agents more reliable.

I’ll close with a live example of an agent exploring a dataset, revising its hypotheses, and choosing what to investigate next.

About the Speaker

Srivatsa P is a member of technical staff at Sigma Computing, where he works on AI agents that reason over complex enterprise data. Previously, he worked on machine learning at Apple and conducted research at the MIT Media Lab; outside of traditional ML, he has also worked on mapping coral reefs through underwater imaging, which sparked an enduring interest in how intelligent systems make sense of messy real-world data.

Does Cross-Validation Work in Deep Learning?

Cross-validation is a standard in hyperparameter optimization in classical ML, but is generally considered computationally prohibitive and unnecessary in deep learning. We investigated whether CV produces 1. better DL models 2. better estimates of future performance of said models. We focused on medical image classification as an important deep learning application domain.

Our results show that cross-validation consistently and materially reduces the gap between development and independent testing error for relatively small training datasets (up to about 1000 images). The test performance improvement was generally modest. We discuss the implications for deep-learning practitioners, including whether the increased computational cost of cross-validation is justified.

About the Speaker

Ljubomir Buturovic is an AI scientist and biotech executive with extensive experience in clinical diagnostics, including work at Inflammatix and Pathwork Diagnostics. His current research focuses on deep-learning image classification, infectious disease adjudication using LLMs, and development of AI models for helping critically ill patients.