BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Meetup//Meetup Calendar 1.0//EN CALSCALE:GREGORIAN METHOD:PUBLISH NAME:Rome AI, Machine Learning and Computer Vision Meetup X-WR-CALNAME:Rome AI, Machine Learning and Computer Vision Meetup BEGIN:VTIMEZONE TZID:Europe/Rome TZURL:http://tzurl.org/zoneinfo-outlook/Europe/Rome X-LIC-LOCATION:Europe/Rome BEGIN:DAYLIGHT TZOFFSETFROM:+0100 TZOFFSETTO:+0200 TZNAME:CEST DTSTART:19700329T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:+0200 TZOFFSETTO:+0100 TZNAME:CET DTSTART:19701025T030000 RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:event_316758156@meetup.com SEQUENCE:1 DTSTAMP:20261024T135600Z DTSTART;TZID=Europe/Rome:20261029T170000 DTEND;TZID=Europe/Rome:20261029T190000 SUMMARY:Oct 29 - How to Build Vision Data Agents with Tools\, Skills\, and MCP DESCRIPTION:Rome AI\, Machine Learning and Computer Vision Meetup\nIn this session\, you’ll learn how to build production-ready AI agents that can reason over your data\, automate complex tasks\, and integrate seamlessly into your existing stack using tools\, skills\, and the Model Context Pro tocol (MCP).\n\n**Date\, Time and Location**\n\nOct 29\, 2026\n9:00 AM - 1 0:00 AM PST\nOnline. **[Register for the Zoom!](https://voxel51.com/events /how-to-build-vision-data-agents-with-tools-skills-and-mcp-october-29-2026 )**\n\nWe’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’l l see how these agents can tag data\, run inference\, evaluate performance \, and surface insights automatically\, all within a unified workflow.\n\n By combining natural language interfaces with programmable building blocks \, teams can dramatically reduce manual effort\, accelerate experimentatio n\, and unlock faster decision-making across the ML lifecycle.\n\nWhether you're building data-centric AI systems\, managing large-scale vision data sets\, or exploring agentic workflows for the first time\, this session wi ll give you a practical blueprint for getting started. URL;VALUE=URI:https://www.meetup.com/rome-ai-machine-learning-and-computer -vision-meetup/events/316758156/ STATUS:CONFIRMED CREATED:20260928T231501Z LAST-MODIFIED:20260928T231501Z CLASS:PUBLIC END:VEVENT BEGIN:VEVENT UID:event_316768251@meetup.com SEQUENCE:1 DTSTAMP:20261024T135600Z DTSTART;TZID=Europe/Rome:20261105T180000 DTEND;TZID=Europe/Rome:20261105T200000 SUMMARY:Nov 5 - Visual AI Agriculture Meetup DESCRIPTION:Rome AI\, Machine Learning and Computer Vision Meetup\nJoin ou r virtual meetup to hear talks from experts on cutting-edge topics at the intersection of agriculture and AI.\n\n**Date\, Time and Location**\n\nNov 05\, 2026\n9:00 AM - 11:00 AM PST\nOnline. **[Register for the Zoom!](htt ps://voxel51.com/events/visual-ai-agriculture-meetup-november-5-2026)**\n\ nTalks will include:\n\n**U-Net Framework for Micro-Scale Surface Damage S egmentation in High-Resolution Soybean Seed Imagery**\n\nAccurate detectio n of surface-level seed damage is critical for soybean seed quality assura nce\, yet automated micro-scale damage detection in high-resolution imager y remains an open challenge due to extreme spatial variability in damage s cale\, 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 surfac e damage\, in 6048 × 4024 pixel images where target damages can be as sma ll as 18 × 18 pixels\, using an augmented dataset of 77\,000 individual s eed images.\n\nTo overcome the difficulties of micro-scale detection\, spa tial sparsity\, and class confusion\, we propose a two-stage training and dual model inference framework built on an optimized U-Net architecture wi th a ResNet34 encoder. In the first stage\, a damage specialist model is t rained using weighted loss functions and a class-balanced approach that pr ioritizes damage classes.\n\nIn the second stage\, transfer learning is ap plied to initialize a healthy seed specialist model from Stage 1 weights\, with rebalanced class weights and tile probabilities that identify health y seeds. During inference\, a confidence-gated damage filter suppresses lo w-confidence predictions\, and healthy seed labels are assigned only when the specialist model's confidence exceeds that of the damage model.\n\nThe two specialist models achieve validation accuracies of 94% and 98.53%\, r espectively\, and the combined inference system successfully detects and l ocalizes all three damage categories across unseen test images under condi tions of extreme spatial variability and class imbalance. Qualitative eval uation confirms close alignment of predicted boundaries with ground truth annotations across all damage types\, including dark spots.\n\nThese resul ts demonstrate that confidence-gated dual model inference can reduce class confusion in imbalanced micro-scale segmentation\, advancing the feasibil ity of fine-grained automated seed quality inspection at the scale.\n\n*Ab out the Speaker*\n\n[Saurav Upadhyaya](https://www.linkedin.com/in/sauravu padhyaya/) is an AI/ML researcher with a Master's degree in Computer Scien ce\, whose work spans agricultural AI\, conversational systems\, and publi c health\, consistently translating advanced technology into tangible real -world impact.\n\n**From camera to drone: wildfire detection and integrati on at the edge**\n\nWildfires are getting larger and more expensive every year. But caught early\, a fire is a small job for a small team\; caught l ate\, 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. \n\nThis talk walks the whole chain: a small vision model running on the c amera that flags smoke\, an integration layer that scores the alert agains t terrain\, weather and history to decide whether it matters\, and a drone dispatched to confirm before anyone commits a crew. We will cover what ha s 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 a t yet.\n\n*About the Speaker*\n\n[Maxime Carriere](https://voxel51.com/eve nts/www.linkedin.com/in/maxime-carriere-871756161) is co-founder of Kernwe rk in Berlin\, where he works on compressing AI models to run on cheap emb edded hardware.\n\n**5\,000 Flights to Answer One Question: Can a Drone Sa mple as Well as a Human?**\n\nEnvironmental testing still begins with a pe rson walking onto a site with a shovel. We built a system that turns a pla in-language brief into a sampling campaign\, flies it\, and returns soil\, water and vegetation samples with GPS\, timestamps and chain of custody i ntact: 5\,000 flights\, roughly 9\,000 miles and 570 flight hours over a 1 85-acre site in Roatan\, much of it under canopy where GPS degrades and cl earances are tight.\n\nI will walk through the vision and planning stack\, including multispectral and thermal canopy sensing\, terrain and land-cov er mapping\, and the constraint solver that clears a route against airspac e and battery budget before anything leaves the ground. Then the harder pr oblem: evidentiary comparability.\n\nA drone-collected result means nothin g to a regulator unless it agrees with a hand-collected split under the sa me method\, so I will show the dual-collection protocol we designed to tes t exactly that. I will close on why calibrated uncertainty and knowing whe n a model should refuse to answer matter more than raw accuracy once the o utput lands in a regulatory filing.\n\n*About the Speaker*\n\n[Rohan Talwa dia](https://www.linkedin.com/in/rohantalwadia/) is co-founder and CTO of LabGPT\, which automates environmental field sampling using drones and edg e analysis\, with results returned through ISO/IEC 17025-accredited partne r laboratories. URL;VALUE=URI:https://www.meetup.com/rome-ai-machine-learning-and-computer -vision-meetup/events/316768251/ STATUS:CONFIRMED CREATED:20260929T183728Z LAST-MODIFIED:20260929T183728Z CLASS:PUBLIC END:VEVENT BEGIN:VEVENT UID:event_317079243@meetup.com SEQUENCE:1 DTSTAMP:20261024T135600Z DTSTART;TZID=Europe/Rome:20261109T180000 DTEND;TZID=Europe/Rome:20261109T200000 SUMMARY:Nov 9 - Agentic Labeling for Specialized Visual Data DESCRIPTION:Rome AI\, Machine Learning and Computer Vision Meetup\n**Date\ , Time and Location**\n\nNov 09\, 2026\n9:00 AM - 10:00 AM PST\nOnline. ** [Register for the Zoom!](https://voxel51.com/events/when-prompting-beats-t raining-agentic-labeling-for-specialized-visual-data-november-9-2026)**\n\ nFrontier multimodal models handle familiar concepts well and struggle on niche ones: sorting drill-core samples into 35 geological classes\, or tel ling adhesion from scarring on a ceramic substrate. Our study across 11 sp ecialized visual datasets in manufacturing inspection\, geology\, and medi cal imaging found that a few labeled examples in the prompt can close much of that gap\, and it also found where that stops working.\n\n**What you ’ll learn:**\n\n* Niche versus hard: how in-context examples supply doma in knowledge a model is missing\, and how to tell a niche task from a genu inely hard one\n* What the study measured: on an industrial defect task\, one labeled example per class raised macro-F1 from 34% to 59% with no fine -tuning\, while on other tasks trained specialists pulled ahead at 5 to 10 examples per class\n* Building a few-shot labeling task: define classes\, pick reference examples\, preview predictions on a small batch\, and refi ne the instructions\n* Finding the crossover point: when to stop prompting a foundation model and train a specialist on bootstrapped labels\n* Runni ng it on your own data: attendees get free GPU compute from Nebius to try the workflow after the session URL;VALUE=URI:https://www.meetup.com/rome-ai-machine-learning-and-computer -vision-meetup/events/317079243/ STATUS:CONFIRMED CREATED:20261021T204232Z LAST-MODIFIED:20261021T204232Z CLASS:PUBLIC END:VEVENT BEGIN:VEVENT UID:event_316572569@meetup.com SEQUENCE:1 DTSTAMP:20261024T135600Z DTSTART;TZID=Europe/Rome:20261111T180000 DTEND;TZID=Europe/Rome:20261111T200000 SUMMARY:Nov 11 - AI\, ML and Computer Vision Meetup DESCRIPTION:Rome AI\, Machine Learning and Computer Vision Meetup\nJoin ou r virtual meetup to hear talks from experts on cutting-edge topics across AI\, ML\, and computer vision.\n\n**Time\, Date and Location**\n\nNov 11\, 2026\n9:00 AM - 11:00 AM PST\nOnline. **[Register for the Zoom!](https:// voxel51.com/events/ai-ml-and-computer-vision-meetup-november-11-2026)**\n\ n**Agentic RAG: Beyond Retrieve-and-Generate**\n\nRetrieval-Augmented Gene ration (RAG) has become the default pattern for grounding large language m odels in external knowledge\, but most implementations still follow a rigi d retrieve-once\, generate-once pipeline — one that struggles with multi -hop questions\, ambiguous queries\, and knowing when its own retrieved co ntext is insufficient. This talk introduces Agentic RAG\, where retrieval is treated as an action within an agent's reasoning loop rather than a fix ed upstream step.\n\nWe'll examine query decomposition and routing for bre aking complex questions into targeted sub-queries\, self-reflective and co rrective retrieval loops that let an agent judge and re-query its own resu lts\, and tool-orchestration patterns (via MCP) that let retrieval sit alo ngside other agent actions like database lookups and API calls. Using a li ve 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 pipelin e still wins.\n\nAttendees will leave with a practical framework for decid ing when the added complexity of agentic RAG is worth it\, and a set of de sign patterns for building it correctly.\n\n*About the Speaker*\n\n[Balaji Venkatasubramaniyar](https://www.linkedin.com/in/balaji-venkatasubramaniy ar/) is a Technical Lead at Wisdom Infotech\, leading a 15+ person enginee ring 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.\n\n**GeoAI for the Physical World: Earth Observation\, Foundation Models\, and Urban Digital Twins**\n\nEarth observation provides a unique form of computer vision for understanding the physical world at city to continental scales. In this t alk\, I will show how satellite imagery\, geospatial data\, machine learni ng\, and foundation-model representations can be combined to characterize urban environments and environmental conditions.\n\nI will present UrbanSc ope Open\, an open GeoAI digital-twin prototype integrating Earth observat ion with 3D buildings\, vegetation\, land-surface temperature\, air qualit y\, noise\, population\, and other urban data. I will also share lessons f rom my research using geospatial foundation-model embeddings for environme ntal prediction across Europe.\n\nThe talk will discuss how these approach es can contribute to increasingly multimodal AI systems capable of reasoni ng about real-world environments.\n\n*About the Speaker*\n\n[Cesar Alvarez ](https://voxel51.com/events/cesar.alvarez@uni-a.de) is a researcher at th e University of Augsburg working at the intersection of GeoAI\, Earth obse rvation\, remote sensing\, and environmental intelligence. His research ap plies machine learning\, computer vision\, and geospatial foundation model s to problems including urban environments\, climate risk\, air quality\, and agriculture.\n\n**Can agents get curious?**\n\nMost AI agents are good at answering a question once we tell them exactly what to look for. The h arder problem is building agents that can explore a complex dataset autono mously: generating hypotheses\, deciding which analyses are worth running\ , allocating additional compute when evidence is ambiguous\, and knowing w hen they have enough evidence to stop.\n\nIn this talk\, I’ll show an ar chitecture for autonomous research agents that combines structured knowled ge\, iterative tool use\, and explicit evidence tracking to turn open-ende d questions into a sequence of testable investigations. I’ll discuss pra ctical 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.\ n\nI’ll close with a live example of an agent exploring a dataset\, revi sing its hypotheses\, and choosing what to investigate next.\n\n*About the Speaker*\n\n[Srivatsa P](https://www.linkedin.com/in/srivatsa-p) is a mem ber of technical staff at Sigma Computing\, where he works on AI agents th at 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 underw ater imaging\, which sparked an enduring interest in how intelligent syste ms make sense of messy real-world data.\n\n**Does Cross-Validation Work in Deep Learning?**\n\nCross-validation is a standard in hyperparameter opti mization in classical ML\, but is generally considered computationally pro hibitive and unnecessary in deep learning. We investigated whether CV prod uces 1. better DL models 2. better estimates of future performance of said models. We focused on medical image classification as an important deep l earning application domain.\n\nOur results show that cross-validation cons istently and materially reduces the gap between development and independen t testing error for relatively small training datasets (up to about 1000 i mages). The test performance improvement was generally modest. We discuss the implications for deep-learning practitioners\, including whether the i ncreased computational cost of cross-validation is justified.\n\n*About th e Speaker*\n\n[Ljubomir Buturovic](https://www.linkedin.com/in/ljubomir-bu turovic-798156/) is an AI scientist and biotech executive with extensive e xperience in clinical diagnostics\, including work at Inflammatix and Path work Diagnostics. His current research focuses on deep-learning image clas sification\, infectious disease adjudication using LLMs\, and development of AI models for helping critically ill patients. URL;VALUE=URI:https://www.meetup.com/rome-ai-machine-learning-and-computer -vision-meetup/events/316572569/ STATUS:CONFIRMED CREATED:20260915T220411Z LAST-MODIFIED:20260915T220411Z CLASS:PUBLIC END:VEVENT BEGIN:VEVENT UID:event_316617439@meetup.com SEQUENCE:1 DTSTAMP:20261024T135600Z DTSTART;TZID=Europe/Rome:20261119T180000 DTEND;TZID=Europe/Rome:20261119T200000 SUMMARY:Nov 19 - Visualizing and Debugging Hugging Face's LeRobot Data DESCRIPTION:Rome AI\, Machine Learning and Computer Vision Meetup\nThis wo rkshop is a hands-on tour of what actually breaks when you load real LeRob ot datasets\, and how Voxel51's native LeRobot support in FiftyOne catches it.\n\n**Date\, Time and Location**\n\nNov 19\, 2026\n9:00 AM - 10:00 AM PST\nOnline. **[Register for the Zoom!](https://voxel51.com/events/lerobot -workshop-november-19-2026)**\n\nRobot learning data doesn't look like the data that trains most vision-language models. It's temporal\, low-dimensi onal in the parts that matter most\, embodiment-specific\, and expensive t o collect\, and increasingly\, it's shipped in Hugging Face's LeRobot form at. But packaging robot data consistently isn't the same as being able to see it: episodes hide broken metadata\, silently dropped sensor modalities \, and mismatched schemas between robots that look identical on paper.\n\n Starting from the core differences that make robot data hard (temporal str ucture\, proprioception\, action grounding\, embodiment heterogeneity\, da ta scarcity)\, we'll move into live debugging: a metadata bug that broke m ultiple dataset imports\, sensor streams that vanish without an error\, an d two robots running an identical benchmark whose data is structurally inc ompatible. By the end\, you'll know what to check before you trust a LeRob ot dataset\, and how to actually look inside one. URL;VALUE=URI:https://www.meetup.com/rome-ai-machine-learning-and-computer -vision-meetup/events/316617439/ STATUS:CONFIRMED CREATED:20260918T224859Z LAST-MODIFIED:20260918T224859Z CLASS:PUBLIC END:VEVENT BEGIN:VEVENT UID:event_316757862@meetup.com SEQUENCE:1 DTSTAMP:20261024T135600Z DTSTART;TZID=Europe/Rome:20261210T180000 DTEND;TZID=Europe/Rome:20261210T200000 SUMMARY:Dec 10 - AI\, ML and Computer Vision Meetup DESCRIPTION:Rome AI\, Machine Learning and Computer Vision Meetup\nJoin ou r virtual meetup to hear talks from experts on cutting-edge topics across AI\, ML\, and computer vision.\n\n**Date\, Time and Location**\n\nDec 10\, 2026\n9:00 AM - 11:00 AM PST\nOnline. **[Register for the Zoom!](https:// voxel51.com/events/ai-ml-and-computer-vision-meetup-december-10-2026)**\n\ n**Your mAP is fine. Your slices are not. Finding the images that actually matter**\n\nA single mAP can look acceptable while one slice is unusable\ , small objects\, occlusion\, or a confused class. I will run a pretrained detector on a few hundred public images (COCO-style\, not a production cr awl)\, report the headline metric\, then break the same run by slice and w alk through the worst samples.\n\nThe claim is modest: you do not need a m illion images or a trained-from-scratch model to see what the mean is hidi ng\, and you cannot relabel everything\, so the next labeling budget shoul d follow those slices. No employer data and no confidential dashboards.\n\ n*About the Speaker*\n\n[Amandeep Jiddewar](https://www.linkedin.com/in/am anfj/) is a Machine Learning Engineer at Pinterest\, with prior work in U. S. manufacturing analytics (CertainTeed / Saint-Gobain) and marketplace al location (QuinStreet). Emory Goizueta MSBA (2019)\, Kaggle Expert. This ta lk is in his personal capacity and does not represent his current or prior employers.\n\n**Production AI Agents: Tool Use\, Evaluation\, Guardrails and Observability**\n\nBuilding an AI agent is easy. Knowing whether it is behaving reliably in production is much harder. This session explores how to evaluate agents that use tools and external systems\, including tool-c all validation\, tracing\, observability\, failure analysis\, guardrails\, permissions\, and human review.\n\nAttendees will learn practical pattern s for understanding not just what an agent answered\, but why it acted and whether the action was safe.\n\n*About the Speaker*\n\n[Akshay Talathi](h ttps://www.linkedin.com/in/akshaytalathi/) is a Vice President at Goldman Sachs with over 10 years of experience building large-scale\, cloud-native \, and distributed systems across financial services and technology. His c urrent focus is GenAI and agentic AI\, including production AI agents\, MC P\, enterprise integrations\, guardrails\, observability\, and reliable AI -driven workflows.\n\n**From Virtual Worlds to Physical AI: Building Simul ation Pipelines That Survive the Real World**\n\nRobots are moving beyond perception systems that only detect and classify objects toward systems th at must understand a scene\, learn behaviors\, and act in the physical wor ld. In this talk\, I’ll show a practical Physical AI workflow using Open USD\, BowerBot\, NVIDIA Isaac Sim and Isaac Lab to build a robot’s virtu al world\, including its body\, cameras\, sensors\, and environment.\n\nWe ’ll explore how demonstrations and reinforcement learning can teach beha viors in simulation\, how vision and sensor data become observations for a robot policy\, and what needs to match when transferring that behavior to real hardware. The demo follows a small robot from a simulated environmen t toward a real-world task\, exposing both the power and the limitations o f sim-to-real.\n\nThe goal is to make the path from pixels to actions conc rete for computer vision and machine learning practitioners.\n\n*About the Speaker*\n\n[Arturo Morales Rangel](https://www.linkedin.com/in/arturo-mo rales-rangel) is the founder of Binary Core LLC\, a Physical AI engineerin g company working across simulation\, perception\, robot learning\, and de ployment.\n\n**Prompting Video-Language Models in Production: Lessons from Enterprise Video Tagging and Evaluation**\n\nVideo-language models can no w describe\, tag and answer questions about long-form video\, but producti on results depend heavily on prompt design and evaluation. This talk share s field lessons from deploying such models for media\, news\, streaming an d sports organizations\, including how to structure prompts for consistent \, schema-bound outputs across hours of footage.\n\nOn one sports video-ta gging project\, consolidating the prompt definitions cut estimated process ing cost about 3x\, and about 90% of the originally reported issues were l ater resolved through prompting alone. I'll also walk through building eva luation frameworks for video outputs\, including for hard domains like ego centric video\, and what changes when migrating between model versions.\n\ nAttendees will leave with reusable prompt patterns and an evaluation chec klist they can apply to their own video pipelines.\n\n*About the Speaker*\ n\n[Yash Gugale](https://www.linkedin.com/in/yashgugale/) is a Senior Forw ard Deployed Engineer at TwelveLabs in New York\, integrating video founda tion models (Marengo for embeddings and search\, Pegasus for video-languag e understanding) for media\, news\, streaming and sports customers. He pre viously spent about six years in video streaming engineering at Peloton an d Samsung TV Plus\, and holds an MS in Computer Graphics Technology from P urdue. URL;VALUE=URI:https://www.meetup.com/rome-ai-machine-learning-and-computer -vision-meetup/events/316757862/ STATUS:CONFIRMED CREATED:20260928T230349Z LAST-MODIFIED:20260928T230349Z CLASS:PUBLIC END:VEVENT X-ORIGINAL-URL:https://www.meetup.com/rome-ai-machine-learning-and-compute r-vision-meetup/events/ical/ X-WR-CALNAME:Rome AI\, Machine Learning and Computer Vision Meetup END:VCALENDAR