Loci

“One of the more interesting pushes to prod.” Last September, two big things happened in one week. The first: CTO Jack Davis and our Machine Learning team were in Milan for the European Conference on Computer Vision. The second: our latest AI models were ready to push to production. A push to production is the moment code enters the live product. It can be nerve-wracking for any developer. Best done in a serene, familiar space, perhaps with a clean desk and multiple monitors. But Jack ships on time. Even if he’s 600 miles away from his desk in London. Even if he’s on a teeming city street by a Milanese canal. And even if he’s equipped with nothing but a laptop, cafe Wi-Fi, and a late night flat white. Five months later, we can confirm the push went flawlessly. The power of a determined CTO (and a really good flat white).

It's often said that you can build a product that’s fast, accurate, or cheap — but you can only pick two. Loci shoots for all three. Our ML Ops team has spent the last several months refining our LLM deployment pipeline to balance speed, precision, and cost simultaneously. The key ingredient is Auto Scaling. Deploy more GPU instances when traffic spikes, and remove them when things quiet down. In his article “Monitoring, Auto Scaling and Load Testing vLLM on ECS,” ML Ops Engineer Hatem Elseidy explains how we built a scaling solution for Loci LLM that saves resources without skimping on performance. It’s a breezy, concise sequel to his first article, "Deploying vLLM on ECS with EC2," where Hatem describes how Loci scaled our AWS infrastructure to process millions of 3D assets per week. Link in comments 🌐

Loci quietly hit a big milestone last autumn: tagging 1M+ 3D assets per week with Loci LLM. Deploying a custom LLM at this level of latency—while keeping costs down—required careful wiring behind the scenes. ML Ops Engineer Hatem Elseidy aced the assignment. In his article, "Deploying vLLM on ECS with EC2," Hatem describes the AWS infrastructure that he and the Loci ML team used to run a vLLM inferencer on a GPU. It is a compelling, concise guide for technical audiences who are interested in ML Ops and know their way around AWS. The article is the latest in a series where Loci’s technical staff explain how we approach complex challenges in 3D and AI. It's our way of working with the garage door open. Link in comments 🌐

The joy of using your own products. Our machine learning team spends hours sifting through massive 3D datasets to review the outputs of our AI models. To make this easier, our resident creative developer Alex Taylor built an internal tool called the Asset Navigator: a speedy browser for viewing 3D assets side-by-side with all the data that our AI models generate about them. We created Asset Navigator to help Loci engineers work faster, but it runs on technology that we offer our customers—like Visualize, Render, Caption, and Embed. Think of it like eating your own cooking. At Loci, we make thousands of API calls to our own endpoints in order to search and understand the 3D content we work with. It's one of the reasons our team is so obsessed with improving what we build. Because when our products are fast, precise, and performant, so are we.

We're delighted to announce Philipp Wulff is joining Loci as a Machine Learning Research Engineer, where he will help build advanced computer vision models for 3D understanding. Philipp graduated with a Master of Science in Robotics, Cognition, and Intelligence from the Technical University of Munich. His research included 3D representation learning, depth-guided dynamic NeRF, and monocular 3D reconstruction. Outside the classroom, Philipp has worked on industrial robots at Alphabet Inc.'s Intrinsic, self-driving cars at Cruise, and traffic simulations at AUDI AG. (His side projects are just as fun; ask him about the summer he built a VR telepresence robot.) Philipp moved to London to work with us and we couldn't be luckier. Send him your local recs—he's especially keen on hackathons. Welcome aboard, Philipp!

Meet Loci Recommend: a discovery engine that lets you use one 3D asset to quickly find more like it. Half the battle of designing a 3D scene is finding the right assets. Good search makes all the difference. Recommend's AI understands art style, content, and context. Which means designers can curate cohesive, style-matched characters and props for a scene in seconds. Below, Recommend examines a 3D scan of a Japanese shrine and suggests similar content from the Quixel Megascans library. We’re making sure that Recommend is easy to customize and plug into existing creative pipelines and marketplace UIs. Adjust thresholds, power a search engine, or create carousels like this one. The possibilities are yours for the making.

“Creating machines that understand the physical world requires teaching them to...understand not just text and images, but also 3D objects,” writes Loci Machine Learning Research Scientist Akshay Gupta. Check out Akshay’s article on how the Loci ML Research team is teaching AI to see and think in three dimensions: https://lnkd.in/eHMV7E9K (Video: Scan of a Triceratops horridus fossil collected by the Smithsonian Institution in 1899, animated by French 3D artist @zacxophone in 2020)

We’re thrilled to welcome Hatem Elseidy to Loci as a Senior Software Engineer, where he spearheads Machine Learning Ops to keep our tools fast and smart at scale. Hatem has spent over a decade building high-impact AI, from ML engineering at Amazon, to computer vision at Synativ, to NLP research at Microsoft Advanced Technology Labs. Outside of being a world-class engineer, Hatem creates 3D short films in Blender, plays and watches football, and is our office sage for video game recommendations. (Ask him for his spreadsheet of all the games he’s completed since 2014.) We couldn’t be luckier. Welcome aboard, Hatem 🚀

Loci just won the Amazon Web Services (AWS) Scaled Thought Award at Dimensions 2024. We couldn’t be more proud. The award recognizes our work on Scene Intelligence, a 3D computer vision model for architects that analyzes and labels every object in a SketchUp design. So architects, contractors, and cost estimators understand exactly what they’re building—down to the last door. Granular data like this makes it easier to create accurate Bills of Materials, estimate costs, audit safety-critical features like fire doors, mark load-bearing elements for structural engineers, and more. Collecting this data used to take human experts hours of manual labeling. Scene Intelligence does it in seconds. It’s part of our quest to build specific, practical AI—the kind that runs behind the scenes to automate the most tedious, error-prone parts of human work. Thank you so much to the good people at AWS and Trimble Inc. for all your support on this journey. And congratulations to our fellow 0-60 Challenge finalists and winners—Foveate, Modelizer AI, ScaffPlan , Skema.AI, Spacely AI, Surfaice, V-Quest, ZIPPZAPP. Pictured: Eugene Yi, Jack Davis, Brett Mitchell

🎙️ Connected in 3D: Real-Time Solutions & Digital Twins in LIVE!!! 🎙️ I've met a lot of incredible people over the years working in 3D and XR so I wanted to share their knowledge with the world. If you're passionate about how immersive 3D tech is shaping the future, get ready for in-depth conversations with the visionaries who are pushing the boundaries of 3D technology, augmented reality, virtual reality, and digital twins to transform industries. From revolutionizing training and maintenance through digital twins to enhancing customer experiences with interactive 3D, each episode dives into real-time solutions that are changing how businesses operate. ✨ Meet Our First Guests: - Albhy Galuten: A Grammy-winning producer, futurist, and pioneering technologist. Known for his work on *Saturday Night Fever* and as a tech leader at Universal Music Group and Sony, Albhy brings incredible insights into innovation and media tech. - Marc Petit: A key player in XR and real-time technology. Formerly head of Autodesk's Media & Entertainment unit and VP and GM of Unreal Engine at Epic Games. Marc is now an independent consultant and investor in the XR space. - Eugene Yi: An accomplished AI researcher and 3D technology innovator with academic roots at the MIT Media Lab and Oxford. His experience in 3D started from Kuku Studios, the animation studio that produced Go! Go! Cory Carson and In Your Dreams. At Loci, he has been pioneering advanced 3D understanding tools for creative studios, game developers, architects, and digital world builders. Stay tuned, subscribe, and be part of the 3D revolution! https://lnkd.in/gdtS35SM #DigitalTwins #RealTimeSolutions #FutureOfWork