Hey! I’m a Research Engineer working on generative AI and multimodal models.
A bit about me:
• 4+ years in ML/CV R&D
• Previously worked on large-scale VLMs at Sber
• YSDA alum
• Published research on fingerspelling recognition at an EMNLP workshop
Currently exploring image generation & editing, diffusion models, and model training.
I’ll be sharing:
→ Intuitions behind ML papers and concepts
→ Things I learn while going down ML rabbit holes
→ Experiments, implementations, and technical insights
→ Occasional thoughts on research and engineering
Mostly writing to understand things better and connect with people working on similar problems.
If you’re into ML research, generative models, or building AI systems, welcome!
Building frontier visual AI shouldn't require billions $ in compute and Internet-scale data.
A year after leaving Meta, I’m excited to share the first independent evaluation by Contra Labs of what we've built at @genpeachAI.
Our first foundation image model. Trained from scratch.
Ranked top-1 on human-centered image generation across all axes! Outperformed the large frontier models.
Small team. A fraction of the resources. More to come! 🚀
@contralabs eval report: https://t.co/HEKY4GRH40
Frontier visual AI shouldn't require billions in compute and Internet-scale data.
Our first model was trained from scratch by a small team, using a fraction of the resources typically required.
Check out Contra Labs' evaluation against 8 leading image models for first results. 🍑
Webcam mode is trickier: depth must be estimated for each frame, so it runs at ~25 FPS, with noticeably lower quality than static photos.
It’s definitely not a replacement for generative relighting. The geometry is approximate, original lighting isn’t fully removed, and shadows aren’t always accurate.
Still, it’s pretty cool how far you can get with classical graphics and a bit of ML.
Recently saw an interactive relighting demo by @krea_ai and got inspired to try something similar locally.
I played around with ray tracing and CUDA a while ago, so I got curious: how far can we get without generative image editing, just using estimated geometry and classical rendering?
Everything runs locally on a single RTX 4060 (8 GB), with rendering taking just ~2–4 ms per frame.
Built LightLab over a couple of evenings with a @claudeai 👇
Under the hood:
• Depth Anything V2 Small estimates relative depth
• Sapiens2 predicts detailed face normals
• Depth is unprojected into approximate 3D geometry
• WebGL handles lighting, shadows, and volumetric effects
For static images, geometry is estimated once. After that, moving lights requires no additional neural network inference.
After two years of rejections, revisions, and rebuttals, we finally made it! 🎉
I guess it’s a good reminder that things don’t always work out on the first try. Sometimes you just have to keep going.
Excited to share that two of our papers were accepted to WSLP @emnlpmeeting 2026!
🤟 HandReader: SOTA fingerspelling recognition on ChicagoFSWild / ChicagoFSWild+.
📊 Znaki: An open Russian fingerspelling dataset with 37K+ videos from 68 signers.
Both works are currently available in a combined arXiv preprint:
https://t.co/iOeuuv3IGj
I don’t have a clear end goal.
I just want to rethink what I know about LLMs and distributed systems — and other things I find interesting.
This account is my way of thinking out loud.