👁️ Talk 1: From Pixels to Actions: Building Vision Agents
Wenchi Ma | Staff Scientist, Computer Vision and Machine Learning, Walmart Global Tech - LinkedIn
How do you turn a raw camera feed into an agent that takes action? Wenchi will share how Walmart builds vision agents for retail environments, from object detection and tracking to the reasoning workflows behind real-world use cases like safety hazard recognition.
🔍Talk 2 Title: Detecting TikTok trends at scale
Joe Tam | Cofounder & CTO @ upriver.ai, prev Instagram & Google - LinkedIn
A TikTok trend is thousands of people copying the same idea - a format, a joke, a transformation, a way of filming something. No two videos look alike: different people, settings, camera angles, and captions. Detecting a trend means answering a deceptively simple question at scale: which of these millions of clips are actually the same thing?
🧠 Talk 3: Beating Video Games with Vector Search
John Kupchanko | Qdrant - LinkedIn
What happens when you point a vector database at a video game? This talk walks through building an agent that uses vector search to play and win: embedding screen frames and game state, retrieving the right memory at the right moment, and the engineering decisions that separate an agent that wanders from one that finishes the game.
🕹️ Talk 4: Understanding 3D Worlds Without Rendering Them
Harish Wajjala | Principal Machine Learning Engineer, Roblox - LinkedIn
What if you could understand a 3D game scene without rendering it first? Harish will explore how asset and position data can support scene-level embeddings for questions and classification, sidestepping the render-then-image pipeline. He’ll connect the approach to Roblox Economy use cases, including avatars and in-game transactions.
🎙️Talk 5: LangChain Voice Agents Evaluations
Caroline di VIttorio | LangChain - LinkedIn
What makes a voice agent reliable beyond simply sounding human? This talk explores how to evaluate voice agents across three dimensions: execution, outcomes, and experience. Learn how to assess whether an agent follows instructions and uses tools correctly, achieves its intended goal, and delivers a smooth conversation without awkward pauses, interruptions, or frustrating delays. Discover how to use LangSmith for tracing and evaluating voice agents, identifying failures, and debugging behavior to build voice agents that are ready for production.
🗄️ Talk 6: Multimodal Models for Databases
Joe Meyer | Senior AI/ML Applied Scientist, Generative AI, SAP - LinkedIn
Enterprise data isn't just text. It's numbers, timestamps, and relationships spread across tables. Joe will explore how multimodal models can treat a database as more than rows of strings, drawing on his foundation model research at SAP with multimodal transformers and heterogeneous graph neural networks that combine language, numeric, and time-series signals. Expect a look at how these models are built and evaluated, and where they open up new ways to reason over structured business data.
🛠️ Community Demos
Built something multimodal, from a vision pipeline to a voice agent to a retrieval-based game? We have three 5-minute demo slots. Message the hosts on Luma with a one-line description of what you’d show.
👥 Who Should Join?
Whether you’re working on computer vision, speech and audio, 3D and generative media, AI agents, search and relevance, embeddings, or vector databases, there’s plenty here for you. In town for AGNTCon or MCPCon that week? This is the evening to meet the SF builders. Come for the engineering insights. Stay for the food, drinks, and people.
Space is limited to 70, and registration is approved by the hosts — register early.
🤝Who Are We
The AI Collective is a global non-profit building the human layer for the AI era. We unite 200,000+ leaders, builders, and stakeholders across 100+ forums worldwide to democratize the frontier, build trust, and coordinate how society navigates the rapid acceleration of technological progress.
Code & Coffee is a meetup for developers to make friends and grow, active in over 30 cities. 100% volunteer-led and volunteer-run, with 80,000+ devs and growing.
Superoptimal runs technical meetups and workshops for AI engineers, bringing practitioners together around real production systems.
Upriver is the API that gives AI agents real-time context on emerging trends and breakout topics, so teams act on what’s timely & relevant for target audiences.
