Roman Degtiarev

Senior ML Engineer at Runpod, London

Roman Degtiarev

Fig. 1The name, sampled from noise in 50 denoising steps. Drag across it to add noise back.

One prompt restyles the name, the portrait and the project previews. Keys 1 to 6 work too.

I make generative models fast, cheap and reliable enough for millions of people to use every day.

Co-founder of Glam AI, which grew from $0 to $50M in annual recurring revenue and 20M+ downloads. Named inventor on three granted US patents for text-guided photo and video stylisation. Now building AI infrastructure at Runpod.

Get in touch Read the story
Roman Degtiarev on stage giving his keynote at Adapty China Conf in Beijing
Fig. 2Keynote at Adapty China Conf, Beijing, August 2026, redrawn in the current style. Hover or tap to see the original. Photo: Adapty.
  • $0 → $50M ARR at Glam AI
  • 20M+ downloads
  • 3 granted US patents
  • Keynote, Adapty China Conf
  • 1st place, Rork Games Hackathon
  • Senior ML Engineer, Runpod
  • PicsArt AI research
  • Lomonosov Moscow State University

About

A bio for organisers and editors. Copy it as it is.

Roman Degtiarev is a machine learning engineer and founder specialising in generative AI and computer vision. He co-founded Glam AI, where he built the technology behind AI photo and video apps that reached more than 20 million downloads and grew from zero to $50 million in annual recurring revenue. He is a named inventor on granted patents in generative AI, computer vision and image and video stylisation, technology deployed in Glam AI's live products.

His work centres on a hard, practical problem: making state-of-the-art generative models fast, affordable and reliable enough for millions of everyday users. Since November 2026 he has been a Senior ML Engineer at Runpod, the AI developer cloud used by more than a million developers. Roman studied at Lomonosov Moscow State University and is based in London.

Based in
London
Now
Senior ML Engineer, Runpod
Before
Co-founder of Glam AI; Mynalabs; PicsArt
Studied
Physics and maths at Lomonosov Moscow State University, then a master's in computer science
Also written
Roman Degtyarev

By the numbers

Glam AI figures as presented in my Adapty China Conf keynote, 2026.

Glam AI annual recurring revenue

$0 → $50M

Grown from zero as co-founder, with unit economics taken from deeply negative to profitable.

Downloads

20M+

iOS and Android, frequently top 10 in the US.

Monthly active users

4M

Using Glam AI's photo and video tools.

Granted US patents

3 + 5

Plus five published international applications.

Engineers in teams I built

40+

Across backend, ML, mobile and DevOps.

The story so far

From satellite images in Moscow to a keynote in Beijing.

  1. Moscow

    Physics, maths, then pixels

    Physics and maths at Lomonosov Moscow State University, then a master's in computer science. Mail.ru's Technosphere programme gave me my first real data pipelines, and my first applied work was detecting vehicles in satellite imagery. Images won me over because the results are visible straight away.

  2. 2019

    PicsArt research

    On PicsArt's international research team I designed a depth-estimation network with four times fewer parameters than MegaDepth at the same quality, extended it to segmentation and made a core warping algorithm 30% faster. Side quests: vector fonts generated from images and 3D photos from a single frame.

  3. 2020

    Mynalabs and a viral accident

    Lead developer on Neiro, an AI video platform with 60+ avatars in 140+ languages. Then I built the real-time face-and-voice technology for an experimental app that let you message friends as someone else. The founders were ready to close it; I took over marketing, analytics and experiments instead, and it went viral, reaching number one in the UAE.

  4. 2023

    The spin-out

    An investor offered to back that product as its own company, with me as co-founder. SaidIt became Molly, and after many iterations Molly became Glam AI.

  5. 2023–2026

    Glam AI

    I owned the technology, backend, product roadmap, hiring, performance marketing and the P&L. We took ARR from $0 to $50M, passed 20 million downloads and 4 million monthly users, shipped product updates three to four times a week and patented the stylisation methods behind the app.

  6. 2026

    London

    Moved to London. Gave the keynote at Adapty China Conf in Beijing, won first place at the Rork Games Hackathon and was part of an award-winning team at the London AI Hackathon.

  7. Now

    Runpod

    Senior ML Engineer at Runpod, the AI developer cloud used by more than a million developers: the layer that decides how far generative AI products can scale.

On stage

Keynote at Adapty China Conf: The Growth Route, Beijing, August 2026.

Roman presenting a slide about plan-picker conversion to the audience at Adapty China Conf in Beijing
Photo: Adapty

From experimentation to scalability: Glam AI growth in practice

6 in 10 tests fail. That is the baseline, not a problem.

23 completed tests in three months, 9 winners, a 40% hit rate.

1

The money metric decides

Conversion is diagnostics, not a criterion. Every test closes on net revenue per user or LTV, with churn, refunds and reviews as guardrails.

2

Winners don't transfer

A mechanic that wins in one place gives zero or worse in another. Channel, placement and segment each need a retest.

3

The pipeline beats the eureka

Design the next iteration before the current one launches. A failed test is knowledge you paid for.

Favourite result

An intro offer on the yearly plan: zero purchases across 2,500 users. We checked. It was not a bug.

What people say

From letters of recommendation, quoted with permission.

“I personally have followed his expert advice repeatedly, and the results have exceeded my wildest expectations.”
Sergey BunasCo-founder of 21st.dev (YC W26), Forbes 30 Under 30 2026
“Roman is the definition of an exceptionally talented founder and engineer. … He is also a very strong expert in computer vision, one of the best I know.”
Victor ShaburovFounder of Looksery (acquired by Snap), inventor on 140+ US patents, investor in Glam AI
“I was eager to bring Roman on board as a cofounder because of his mad computer vision skills and his broader wisdom and experience in getting great products to customers.”
Sergey Bunas21st.dev
“Roman is also a master of experiments and positive unit economics.”
Victor ShaburovLooksery, Snap

Projects

Hover a row for a generated preview.

  • Glam AI

    AI photo and video studio: hundreds of effects, AI photoshoots and virtual try-on. 20M+ downloads, $0 to $50M ARR.

    App StoreGoogle Play
    2023–2026Co-founder

  • Neiro Studio

    AI video with 60+ realistic avatars, custom voices and emotions, translated into 140+ languages, from script to finished video.

    2020–2022Lead developer

  • Friendsheep

    A party game anyone joins by scanning a QR code. During the demo the whole audience played at once. First place at the Rork Games Hackathon.

    Hackathon on Devpost
    2026Hackathon win

  • ChaChat

    AI character chat with more than 50,000 characters, memory and voice conversations.

    2023Mynalabs

  • Monocular depth estimation

    A depth network with four times fewer parameters than MegaDepth at equivalent performance, later adapted to segmentation.

    2019–2020PicsArt

Patents

Methods for steering image and video generation with plain language, built for and deployed in Glam AI's apps.

GrantedUS 12,205,254B1

AI-driven video stylisation with text description tuning

Restyles an input video from a text prompt: frames are warped, encoded into latent space and regenerated under the guidance of a natural-language description.

View on Google Patents
GrantedUS 12,211,180B1

AI-driven photo stylisation with text description tuning

Text-guided image stylisation with latent diffusion: the photo and the prompt are encoded into a shared embedding space to generate the styled result.

View on Google Patents
GrantedUS 12,288,280B1

Scene, pose and environment replacement in videos and images

Generates a new scene from three inputs: a face image, a body-pose reference and a text description of the target environment.

View on Google Patents

Published international applications

Playground

A pocket-sized cousin of the patents above. It runs entirely in your browser, so your photo never leaves your device.

Fig. 3Test scene in halftone. Drop a photo here to try your own.
Picture

Hackathons

Six awards, built under deadline.

2026, London

Rork Games Hackathon

First place

Friendsheep: the whole audience joined by QR code and played live during the demo.

Play itDevpost

October 2026, London

London AI Hackathon

Award-winning team

Built with a team that included Imperial College AI students.

2018, Barcelona

HackUPC

Winner among 329 participants

HealthChain, a working prototype built in 36 hours at one of Europe's largest student hackathons.

Devpost

2018

SmartMail Hack 2 (Mail.Ru)

Special jury award

Team OneLayerNetwork: machine learning for document search and classification inside email.

2018, St Petersburg

Photo Lab Hackathon

Winner

A web service that turns user photos into shareable visuals, built against the clock.

Snapchat

Snapchat Photo Hack

Winner

Photo and AR experiences for Snapchat's platform.

Bring me in to speak, judge or teach.

Based in London and open to events across the UK and Europe: talks, juries, guest lectures and mentoring.

Talk topics

  1. Taking a generative AI app from $0 to $50M ARR
  2. Patenting style-transfer technology inside a consumer app
  3. Why consumer AI is now an infrastructure problem
  • Talks and podcastsShipping generative models to millions of users: latency, cost and what people keep using
  • JudgingHackathons and awards in AI, computer vision and video generation
  • Guest lecturesUniversity AI societies and courses on generative models in production
  • MentoringFounders and engineers building consumer AI products
LinkedIn GitHub Devpost