She spent years at the ballet barre learning that the effortless takes ten thousand repetitions. Now she is teaching a single photograph to look up and talk back.
Co-Founder & CEO, LemonSlice · San Francisco
The face behind the faces. Colucci builds models that make video answer back.
Dateline: San Francisco
Upload one image. A corporate headshot, a cartoon animal, a Renaissance painting. Press start, and the thing looks at you and begins to speak - in real time, with hand gestures and a turn of the head, no custom training, no waiting. That is LemonSlice-2, and the person who decided this should exist is Lina Colucci.
Colucci is the co-founder and CEO of LemonSlice, a frontier AI lab she runs out of San Francisco with co-founders Sidney Primas and Andrew Weitz. In December 2025 the company came out of stealth with $10.5 million in seed funding and a claim it is happy to defend: the world's first interactive talking AI video model. Where most generative video produces a clip you watch, LemonSlice produces a face you converse with. The difference, Colucci argues, is the same one large language models drew between a printed page and a chat window.
Her pitch is not subtle, and she does not want it to be. "In the future, all video will be interactive and personalized to whoever is watching," she says. "We're building the technology that makes that possible." It is the kind of sentence that sounds either obvious or absurd depending on the decade you say it in. Colucci is betting on obvious.
The difference between impossible and inevitable? That's about a thousand training runs you haven't done yet.
- LINA COLUCCI
By The Numbers
The Work Now
The clearest way to understand LemonSlice is to picture every voice agent that already exists - the support bot, the AI tutor, the assistant on the other end of a phone tree - and then give it a face that reacts while it speaks. LemonSlice ships as an API for developers and an embeddable widget for websites. Feed it audio, hand it an image, and it returns a real-time avatar that nods, gestures, and holds eye contact.
The technical bet underneath is harder than the demo lets on. Real-time interactive video means the model cannot pre-render and polish; it has to generate believable motion as the conversation happens, fast enough that the lag never breaks the spell. The investor list suggests people who have built hard infrastructure take the bet seriously: Matrix Partners and Y Combinator led, joined by Arash Ferdowsi, the co-founder and CTO of Dropbox, and Emmett Shear, the co-founder and former CEO of Twitch. The EDM-pop duo The Chainsmokers are on the cap table too - a reminder that Colucci is building for entertainment as much as enterprise.
LemonSlice grew out of Infinity AI, the company Colucci took through Y Combinator's Winter 2024 batch. Infinity started as a synthetic-data company - generating artificial training data so Fortune 100 firms could build AI models without harvesting real-world footage - and evolved toward the video foundation models that became LemonSlice. The soft launch to her YC batch left a mark. "I am blown away," she wrote the morning after. The thread that runs through all of it is a single conviction: video is about to stop being something you watch and start being something that watches back.
Zero-shot, real-time avatars from a single image. Headshots, cartoons, paintings - all fair game. Full face, hands, and body motion.
Infrastructure other companies build on - a face layer that voice agents and apps can plug into directly.
The Bet
There is a reason Colucci keeps reaching for the language-model comparison. For decades, text on a screen was something you read top to bottom and then closed. Chat changed the grammar of it: now text answers, remembers, and adapts to the person typing. Her wager is that video is sitting exactly where text sat before that shift - rich, expensive to produce, and fundamentally one-directional. You press play, it performs, you leave.
LemonSlice's product is an argument that the one-directional era is ending. If a video can be generated on the fly and personalized to whoever is watching, then the unit of media stops being the finished clip and becomes the live exchange. A teacher who never tires. A character from a film who steps out of the frame to answer a child's questions. A brand spokesperson who speaks every customer's language and remembers the last conversation. None of that works if the face is pre-recorded. All of it works if the face is generated in real time - which is the specific, unglamorous engineering problem Colucci's lab decided to own.
The choice to ship as infrastructure rather than a single flashy app is a tell about how she thinks. LemonSlice positions itself as the layer other companies build on, the same way a payments company would rather be the rails than one storefront. It is a quieter ambition on the surface and a larger one underneath. The co-founding team - Colucci alongside Sidney Primas and Andrew Weitz - carries degrees from MIT, Harvard, Stanford, and Duke between them, the kind of bench you assemble when you intend to solve the hard version of a problem instead of the demo version.
Matrix Partners and Y Combinator, with Dropbox's Arash Ferdowsi and Twitch's Emmett Shear - operators who have shipped hard infrastructure themselves.
The EDM-pop duo on the cap table is a hint: this is built for entertainment and culture as much as for the enterprise stack.
The Barre
Before the PhD and the cap table, there was the studio. Colucci trained as a professional ballerina, and she will tell you it taught her more about building than any business book.
"Dance is a peak life experience where my body, heart, and mind all align," she says. What an audience reads as effortless on stage is the visible tip of thousands of unglamorous hours at the barre - the same arithmetic, she points out, as a model that finally works after a thousand training runs nobody applauded.
It bleeds into how she launches products. Colucci treats a launch like a performance: something built to entertain first, so the demonstration of the product arrives as a kind of magic trick rather than a slide. She compares the daily grind of AI research to being a 16th-century explorer charting territory no map has yet drawn. The romance is sincere. So is the discipline underneath it.
Her own summary of the method is plain: "I believe that reaching your full potential takes both working hard and putting yourself out there every day." Two verbs. Both required.
The Range
At seventeen, Colucci co-designed a redesigned ballet pointe shoe with Nike and wrote it up as a first author. The paper landed on the cover of Ergonomics in Design. Most people's teenage projects do not get peer-reviewed; hers set the tone.
Born in Sao Paulo, Brazil, she grew up moving between Brazil, Canada, and the United States, with research stints in India and Sweden along the way. She studied mechanical engineering at Duke University, graduating with Departmental Distinction in 2012 as one of thirty Robertson Scholars - a full merit scholarship that let her enroll at both Duke and UNC Chapel Hill. A semester at Stockholm's KTH added biomechanics and bionanotechnology to the mix.
Then came the Harvard-MIT Health Sciences and Technology program, where she earned a PhD in medical engineering and medical physics. Her dissertation work produced algorithms for a portable, non-invasive sensor that read the body's hydration state; the research ran as the cover story of Science Translational Medicine. She has authored or co-authored more than seven peer-reviewed papers and patent applications, given a TEDxBrussels talk titled "Why We Should All Hack Medicine," and founded EDGE Analytics, a machine-learning consultancy she now advises as a board member, before any of the AI-video work began.
Graduates Duke in mechanical engineering, with Departmental Distinction, as a Robertson Scholar.
Harvard-MIT HST. Hydration-sensor research becomes the cover of Science Translational Medicine.
Founds Infinity AI - synthetic training data for the Fortune 100. Raises a $5M seed.
Goes through Y Combinator (W24); pivots toward interactive AI video.
LemonSlice debuts publicly: $10.5M seed and the LemonSlice-2 real-time avatar model.
In the future, all video will be interactive and personalized to whoever is watching. We're building the technology that makes that possible.
- LINA COLUCCI, ON THE LEMONSLICE THESIS
Off The Clock
The ballerina is also a clarinetist, comfortable in both classical and jazz registers. She is an avid cyclist and a serious art enthusiast, which goes some way to explaining why a video model that animates Renaissance paintings feels, in her hands, less like a gimmick and more like a thesis.
She also hosts The Edge of Infinity, a podcast where she sits down with AI founders, engineers, and researchers - Coactive AI's Cody Coleman, Symbolica's George Morgan, Stanford's Akshay Chaudhari - to argue about how artificial intelligence reshapes the world in the near term and out toward, well, infinity. True to form, she once built an entire episode as a fully AI-generated podcast, because the most convincing way to make her point was to perform it.
The Directory
Share This Profile