leffff - Overview (original) (raw)

Lev Novitskiy leffff

Block or report leffff

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 5.0 Released 8 models
  • Developed a universal Multistage Video Diffusion Distillation Pipeline (х2 - х8), CFG distill -> Consistency distill -> Adversarial Distillation
  • Kandinsky 4.1 Video Distillation (х8 speedup)
  • WAN 2.1, Hunyuan Video Distillation (х8 speedup)
  • 1000+ GPU distirubuted training pipeline (PyTorch)
  • Kandinsly 4.0 T2V Flash (Distillation x25 speedup)
  • Video Diffusion Distillation (LADD, ReFlow, LCM)
  • Auto-Encoders (VQ, KL) Created SOTA KL-VAE for LDMs
  • FlowModels Inprementation of Flow-based models and distillation technoiques
  • Dragon Diffusion on Kandinsky 3.0
  • 3D Microstructures generation
• 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)

🎓 Education

• Skoltech
• NUST MISiS (Since 2021)

Cources

• Yandex.Lyceum (2018 - 2020) Diploma, Verification
• Deep Learning School (Since 2020 - 2022)
• Tinkoff Generation ML Course (2022 - 2023) | Enrollment Solution
• Tinkoff Generation Advanced DL Course (2023) | Enrollment Solution
• Summer with AIRI 2023 (2023) | Project
• SMILES 2025 | Project

🏆 Competition Background

• 🥇 AIIJC "AI in Customer service" 2021 | Solution Description
• 🥇 HSE Assistant Hack: Python "LLM assistant for programming contests" 2024 | Solution
• 🥇 RuCode 6.5 "AI-generated text classification" 2023 | Solution
• 🥇 ML Talent Match "NER on resumes 2024" | Solution
• 🥈 LCT Yakitiya | Solution
• 🥈 RuCode 7.0 | Solution
• 🥈 Data Wagon Hack 2023 | Solution
• 🥈 RuCode 6.0 "Car color classification" 2022 | Solution
• 🥈 RuCode 5.0 "AI in vocation salary prediction" 2022 | Solution
• 4/80 Moscow Travel Hack "ServiceTech" | Solution
• 4/20 Tula Hack "Building Detection from satellite images" | Solution
• Finalists VTB More Tech 4.0 | Solution
• 24/113 Leaders of Digital "AI in prediction of RBC news popularity" | Solution
• 21/76 Yandex ML Cup NLP | Solution
• 11/50 NTI BD&ML 2020-2021
• 168/501 MTS ML Cup | Solution

🎉 Other Achievements

• Tinkoff Scholarship holder (2022 - 2023)
• VK Education Scholarship holder (2023 - 2024)

👨‍🏫 Other activities

• MISIS AI Lab lecturer
• • Decision Trees, Bagging, Boosting

🐶 Projects

• FlowModels (IN PROGRESS)
• Graphormer in PyTorch-Geometric
• Adversarial Diffusion Distillation implementation
• Diffusion Forcing with Flow Matching
• Dragon Diffusion on Kandinsky 3.0
• 3D Microstructures generation
• Diffusers-like Euler Scheduler for Flow Matching
• South Park Character Generation, Dataset
• Flow Matching on ODEs and SDEs for colving inverse problems (Summer with AIRI 2023)
• Diffusion pet project
• Methods for Pretraining Graph Neural Networks
• Tools for hyperspherical coordinates
• Variational Graph Auto-Encoders
• Point cloud project
• Tools for model ensembling
• My Gradient boosting
• Web app for conducting ACM competitios

Languages

🇷🇺 Russian - Native
🇬🇧 English - C1
🇨🇳 Chinese - A2

Hobbies

• 🏄‍♂️ Surfing
• 🏊‍♂️ Swimming
• 🏂 Snowboarding

Contacts

My blog

Pinned Loading

1. 🤗 Diffusers: State-of-the-art diffusion models for image, video, and audio generation in PyTorch.
   Python 32.1k 6.6k
2. Kandinsky 5.0: A family of diffusion models for Video & Image generation
   Python 577 39
3. Text and image to video generation: Kandinsky 4.0 (2024)
   Python 149 12
4. The aim of this repository is to test and implement Flow-Matching-based models
   Jupyter Notebook 120 3
5. Microsoft Graphormer (https://arxiv.org/abs/2106.05234) rewritten in Pytorch-Geometric
   Python 163 21