LE

Lev Novitskiy

@leffff

Research Engineer @ Kandinsky Core Pretrain, Skoltech DS

Moscow, Russia

Website

Joined November 2019

227 Followers

218 Following

AI Insights & Achievements

42/ 100

Quality

Avg 13.5 stars per repo.

Growth

6 new projects with 72 stars.

Veteran

Coding for 7 years.

Profile README

Lev Novitskiy

Machine learning engineer / Data Scientist

Fields of interest: Diffusion Distillation, Diffusion, Mulimodal models, Flow Matching, GANs

CV

📊 ML stack

🔧 Other skills

👨‍💻 Work experience

DL Researcher at SBER AI (April 2024 - Present day):
- Kandinsky 5 Pro - best T2V open source model on LLM Arena
- Kandinsky 5 Diffusers integration
- Kandinsky 5 Pro and Lite Distillation (x6 Speedup with 45% Human Preference Score)
- Novel step caching strategy for Diffusion models (NDA) with 2.0–3.5x Speedup
- Kandinsky 5 RLHF Pipeline, Open-source project on Diffusion Reward Modeling for Text Rendering
- Kandinsky 4.1 Released, 3 stage Distillaiton Pipeline (x8 Speedup with 51% Human Preference Score)
- WAN 2.1, Hunyuan Video Distillation
- 1000+ GPU training (PyTorch), large pipeline for diffusion pre-training with Tensor Parallel + FSDP
- Stable pipeline for training 40b GAN model. Proposed custom GAN Parallelism (instead of TP) for training Large Diffusion GAN Models
- Created Kandinsky 4.0 T2V Flash (x25 speedup with 45% Human Preference Score)
- CogVideoX Diffusion GAN Finetuning (59% Human Preference Score)
- Image Auto Encoder for Kandinsky (Proposed SOTA VIVAT VAE)
- Dragon Diffusion editing with Kandinsky 3
DL Researcher at AIRI (April 2023 - May 2024)
As a result of the research we propose several new models for crystal structure generation and modification code
- New materials design with VAEs (VAEs)
- Formation Energy regression with neural networks (GNNs, PointNet, CNNs, Transformers)
- Crystall structure generation, optimization (Diffusion, Flow Matching, Bridge Matching, Auto-Encoders, CNNs, Transformers, En-Transformers)
DL Engeneer at SBER Cyber Security (July 2022 - April 2024):
- NLP: fraud call analisys
- Classic ML: fraud detection with gradient boosting
- Transaction Graph Neural Networks (GCN, GAT, GraphSAGE, Graphormer, GIN)
- Transaction Graph Neural Network pretraining (self-supervised: ARGA, ARGVA, GAEs, VGAEs, Contrastive Learning)
- Transaction Graph Neural Network (Temporal GNNs)
- Transaction sequence scoring (LSTM, GRU, Transformer Encoder)
- User scoring (Gradient Boosting, Ensembling)
- Vulnerability detection in assembly with neural networks (Angr, GNNs)
- Antifraud Voice Bot for preventinng call fraud (BERT, NSP, Pretraining, SFT, GPT, FAISS, HNSW, Retrieval, Whisper, ConFormer)