Sergei Bratchikov

- Building scalable infrastructure for training, evaluating, and deploying advanced AI safety systems. - Developing distributed training pipelines for large-scale language model research and moderation models. - Designing model architectures, training workflows, and evaluation systems for robust AI behavior control. - Working on high-throughput data processing, experiment management, and model optimization infrastructure.

- Leading a development team in building OpenSource LLM models and datasets for Russian and English (Vikhr-Nemo, Vikhr-Llama). Models used on many competitions and companies, with many public usage reports. - Developing open-source frameworks for LLM training and evaluation (effective_llm_alignment, ru_llm_arena). - RnD in offline & online Reinforcement Learning for NLP, creation of opensource SMPO and offline GRPO methods - Writing and publishing research papers at conferences (publication at EMNLP 2024). - Development of cross-lingual LOGIC-701 benchmark (cited in works on reasoning LLMs like K2-V2, AM-Thinking-v1, etc) HuggingFace account: https://huggingface.co/Vikhrmodels GitHub account: https://github.com/VikhrModels Important: This is a non-commercial activity.

- LLM alignment for Alice assistant in Russian and other languages (RLHF, DPO, SFT, etc.) - Reward models training based on assessors and synth markup. - Working with datasets, generating specific data, preparing data processing and improvement pipelines (YQL/SQL) - Distributed training of large models on Multi-GPU clusters - Incremental improvements of users' experience with AI system based on business metrics - Speculative decoding methods, training modules for LLM inference acceleration, quantization in FP8, large-scale inference Stack: Python, TensorRT, triton-inference-server, YQL/SQL

Right now my tasks are: - Architecture modification, additional training and LLM inference (7-13B Llama, Mistral) to create a banking assistant service supplemented with bank API tools. - Inference optimizations of large models with vLLM, ONNX, Triton - LLM training in function calling skills, building RAG pipelines and creating scripts for the assistant using them. - Open source release of ruRoPEBert models on HuggingFace with article on Habr - Wide application of LLM for various other scenarios (feedback analysis, document processing) - Training auxiliary NLP pipelines for LLM using DeBERTa - Hiring prompt engineers

My tasks were: - Building a complex dialog system, imitating various interlocutors - Acceleration of text generation and processing pipelines with large language models (GPT-J, BERT, etc). Use of ONNX, DeepSpeed and TensorRT tools for highload inference and training. - Refinement of the functionality of the Optimum Huggingface library for effective GPT text generation

- Semantic analysis of the digital footprint of users (comments, reviews, etc.). - Intelligent analysis of the job market of the largest companies for comparison with educational programs and their automated evaluation. Usage of different types of BERT-like models for summarization, information retrieval, relevance scoring. - Creation of a NLP tool for automated verification of large scanned documents using OCR - Participation in the development of a platform based on OpenEDX - Participation in the creation of technical specifications for the development of various high-load services - Deploying ML models as REST-based services with FastAPI

The main task was the processing of scanned documents, respectively, libraries such as OpenCV, TesseractOCR, transformers and others were used. A number of tasks related to the reconstruction of dialogues and recommendation systems based on NLP were also performed. Several full-fledged RESTful services were developed using the FastAPI and Django framework, and I also got experience working with Airflow.

Participation in the development of an HR platform (user profile) for all employees of the company, using the Spring and SQL framework, writing autotests. Teamwork based on Agile methodology.

Development and testing of robot localization algorithms in an environment using Deep Learning based on panoramic camera images. Also conducting experiments and comparing different SLAM algorithms.