He spent fifteen years teaching machines to recognize a face in a photo, a frame in a video, a cat in a crowd. Now he is teaching them to read a blueprint.
The vision guy, looking back at us
Aconstruction set for a single building can run thousands of pages. Linework, tags, schedules, specifications, revisions stacked on revisions. A wall on sheet A-201 references a detail on S-304 which contradicts a note on M-110. Somewhere in that thicket is a conflict that, if missed, becomes a six-figure mistake poured in concrete.
Lubomir Bourdev's newest company, Primepoint, builds an AI that reads the whole set the way a veteran project engineer would - except it never gets tired and never forgets which sheet said what. It traces the linework, parses the tags, and links every document into one knowledge graph. Ask it a question in plain English and an assistant named Marvin answers.
It is a strange place to find a man who helped build Facebook AI Research. Then again, the through-line of his career is not Facebook or Apple or Adobe. It is a single stubborn obsession: getting computers to understand what they are looking at.
What we are building enables teams to quickly surface discrepancies in drawings earlier.
At Adobe, before "deep learning" was a phrase anyone said at parties, he shipped a face detector built on a Soft Cascade architecture into Photoshop Elements. It was, by most accounts, the first face detection most consumers ever touched. He also co-wrote the Generic Image Library, which still lives inside the Boost C++ libraries, and designed the Transparency Flattener that quietly powers Illustrator, InDesign and Acrobat.
Then UC Berkeley, a PhD under Jitendra Malik, and Poselets - part-based models that learned to separate how a body is posed from how it looks. The work helped seed a generation of person-detection research.
He was the first person Facebook hired for computer vision, and a founding member of Facebook AI Research. With Manohar Paluri he built the object-recognition engine that runs on every photo and every second of video across Facebook and Instagram. Cade Metz put him in the book "Genius Makers," in the room where that lab was born.
In 2015 he left to chase a harder problem: video compression, reimagined as a learning problem instead of a hand-tuned standard. WaveOne's learned codec beat H.264, H.265 and AV1 in low-latency mode. The Wall Street Journal ran a front-page feature calling it "The Real Pied Piper." In 2023, Apple acquired the company.
Primepoint was founded in 2024 with Hamid Palo, an early Trello employee who helped scale that product, as co-founder and CPO. Kamran Azarbal, a decade at Webcor who climbed from field engineer to project director, runs strategy.
The pitch is deceptively plain: most "construction AI" treats a drawing as a flat image. Primepoint treats it as a document with structure - reading linework, tags and cross-references, then connecting drawings to schedules and specs in a knowledge graph. It plugs into Procore and Autodesk Construction Cloud, and it does not train external models on customer data.
The seed round arrived in two tranches: $4M co-led by Penny Jar Capital and NextView Ventures, then $6M led by Navitas Capital, with GS Futures and Aglaé Ventures along for the ride. Early partner Sundt Construction put it to work on a university campus project in Arizona.
Backers: Penny Jar · NextView · Navitas · GS Futures · Aglaé · angel: Yann LeCun
Before NLP was a field anyone funded, he won Bulgaria's national Computational Linguistics competition - first in 1992, second in 1994.
The Wall Street Journal's 2018 front page borrowed HBO's fictional compression startup to describe his very real one.
Adobe, Facebook, WaveOne, Primepoint. Each time he walked away from a working thing to build the next hard thing.
Yann LeCun, who helped found the lab Bourdev worked in, now writes him angel checks.
His Generic Image Library outlived every product it shipped in - it is still part of the Boost C++ libraries.
A MENSA member who practices martial arts and tennis and has stamped a passport in 30-plus countries.
His Adobe Transparency Flattener was licensed out to other companies, Kodak among them - infrastructure most designers never knew they were standing on.
He has served as an Area Chair for CVPR and ECCV, the conferences where the field decides what counts as progress.
His object recognizer has been run more than 500 billion times. That is not a typo. That is the scale of "every photo."
He keeps choosing problems where the model has to understand, not just classify - photos, then video, now the dense grammar of drawings.