Dec 10 - AI, ML and Computer Vision Meetup

The AI, ML and Computer Vision Meetup is an online session on December 10, 2026, from 9:00 AM to 11:00 AM PST, with four expert talks on AI, machine learning and computer vision. It runs on Zoom, and the organiser asks attendees to register for the Zoom link.

• Your mAP is fine. Your slices are not. Finding the images that actually matter by Amandeep Jiddewar, a Machine Learning Engineer at Pinterest. He runs a pretrained detector over a few hundred public COCO-style images and reports the overall metric. He then splits the same results by slice, such as small objects, occlusion and confused classes, and examines the worst samples. His point is that future labeling budgets should target weak slices. The talk uses no employer data and is given in a personal capacity.
• Production AI Agents: Tool Use, Evaluation, Guardrails and Observability by Akshay Talathi, a Vice President at Goldman Sachs. The session covers how to evaluate agents that call tools and outside systems. Topics include tool-call validation, tracing, failure analysis, guardrails, permissions and human review, with the aim of understanding why an agent acted and whether the action was safe.
• From Virtual Worlds to Physical AI: Building Simulation Pipelines That Survive the Real World by Arturo Morales Rangel, founder of Binary Core LLC. He demonstrates a Physical AI workflow built with OpenUSD, BowerBot, NVIDIA Isaac Sim and Isaac Lab. It covers building a robot's simulated body, sensors and environment, teaching behaviors through demonstrations and reinforcement learning, and what has to match when moving to real hardware. A small robot is followed from simulation toward a real task.
• Prompting Video-Language Models in Production: Lessons from Enterprise Video Tagging and Evaluation by Yash Gugale, a Senior Forward Deployed Engineer at TwelveLabs. He draws on deployments for media, news, streaming and sports organizations, including prompt structure for schema-bound output over long footage. On one sports tagging project, merging prompt definitions cut estimated processing cost about 3x, and about 90% of the originally reported issues were later fixed through prompting alone. He also covers evaluation frameworks for video output, including egocentric video, and changes when moving between model versions. Attendees get reusable prompt patterns and an evaluation checklist.