Vision-Language Models in the Real World: What Ships, What Breaks, What’s Next


Date:  Tuesday, May 12

Start Time: 10:25 am

End Time: 11:10 am

Vision-language models are moving fast, but it’s not always clear where they add value, and many teams struggle to turn demos into dependable products. In this plenary panel, we’ll cut through the hype and focus on where VLMs make sense and the barriers to deploying them in real systems. We’ll discuss where VLMs are delivering clear value and where classic CV still wins on cost, latency and determinism. Panelists will compare practical hybrid architectures, examine failure modes such as weak grounding and hallucination and outline guardrails, evaluation methods and monitoring that work in production. We’ll also debate the edge vs. cloud split, domain adaptation strategies and the privacy/security governance required when models can answer open-ended questions about people and places. Attendees will leave with insights into requirements and potential pitfalls to de-risk their VLM road map.

Track

Session Speakers

  • David Selinger is a technologist at heart and has founded a number of successful start-ups, including Redfin and RichRelevance (and some unsuccessful ones too!). He has served a number of nonprofits, including My Two Front Teeth (now the Family Giving Tree), the Darfur Stoves Project (now Potential Energy) and Silicon Climate. His favorite technology right now is deep learning, specifically as applied to image and video object detection/recognition. David believes that this technology is undergoing a disruptive improvement and will change numerous industries. That’s why he started Deep Sentinel—which puts deep learning to practical use protecting families and businesses.

  • Phil Lapsley is a Co-Founder and Vice President at the engineering consulting firm BDTI and Vice President of Business Development at the Edge AI and Vision Alliance. At BDTI he leads computer vision and embedded AI projects. Previously, he worked at several start-ups, as well as at McKinsey & Company, a management consulting company that advises Fortune 100 companies on business strategy. He holds BS and MS degrees in Electrical Engineering and Computer Sciences from UC Berkeley and an MBA from the MIT Sloan School of Management. He is the author of Exploding the Phone (2014), a history of the first network hackers, co-author of DSP Processor Fundamentals (1997) and inventor or co-inventor on 26 patents. He is also the co-creator of Network News Transfer Protocol (NNTP; Internet standard RFC 977) and was a contributor to the Berkeley UNIX project.

  • Rakshit Agrawal is a Principal Applied Scientist at Microsoft with deep expertise in machine learning, computer vision and artificial intelligence. He holds a PhD in Computer Science from the University of California, Santa Cruz, where his dissertation focused on generalized learning models for structured data. Rakshit has led AI research and development at start-ups, including Camio and Synthpop AI, and has held research roles at Microsoft and NVIDIA. His work spans AI, agentic systems, computer vision and technology for social good, with a track record of translating cutting-edge research into real-world products.

  • Tushar Sheth is a Co-Founder and the CEO of Superfocus.ai, a seed-stage start-up that builds superhuman AIs that read, write, listen, speak, decide and act. It is backed by Cercano Management, Trust Ventures, Metaplanet and other leading investors. Tushar was previously a political appointee and policy advisor in the Obama White House. He studied engineering and law, both at the University of Michigan, and practiced law at Simpson Thacher & Bartlett and the Asian American Legal Defense and Education Fund in New York.

  • As Co-Founder and CTO of Hayden AI, Vaibhav Ghadiok is architecting the physical AI operating system for cities: a vertically integrated, edge-to-cloud AI platform that transforms urban infrastructure into a persistent learning system. What began as a blank sheet is now a company backed by over $100M from top-tier investors, with over $100M in ARR and deployments across New York, Los Angeles, Chicago, Philadelphia and Washington, DC. Under his leadership, Hayden has built a vertically integrated stack—co-designing sensing hardware, perception, machine learning and cloud infrastructure—optimized from the moment a photon hits the lens to large-scale analytics in the cloud. The platform fuses multimodal sensing, geometric priors and learned behavioral priors to create a persistent 3D world model of complex urban environments. This technical foundation is protected by an IP portfolio of 20+ patents across sensing, perception and system architecture. The impact is measurable: bus collisions reduced by up to 35%, emissions lowered by 10% and transit speeds improved by as much as 40%—affecting more than 100M riders annually in New York alone. Vaibhav leads a 100+ person multidisciplinary engineering organization spanning more than twenty technical domains, aligning frontier research with production execution at scale. His expertise sits at the intersection of 3D perception, robotics and large-scale AI systems—translating deep technical innovation into durable, infrastructure-grade platforms. Earlier, he pioneered visually guided autonomous drones and demonstrated the world’s first instance of autonomous aerial manipulation. Vaibhav believes the long-term impact of AI will come not only from increasingly powerful models, but from systems embedded in the physical world—continuously learning, compounding and improving.


Share

Vision-Language Models in the Real World: What Ships, What Breaks, What’s Next