Skip to content
DSML.KZ
BS

Bogdan Suvernev

@Bogdan

Machine Learning Engineer | LLMs, Reinforcement Learning, RAG & Distributed Training

I am a Computer Science graduate from City University of Hong Kong with hands-on experience in machine learning, large language models, and reinforcement learning.

Featured

About

I have worked as both a Research Assistant and AI Solution Engineer, developing LLM training pipelines, RAG systems, and scalable AI solutions. Passionate about advancing AI research, I enjoy building efficient deep learning systems and applying cutting-edge techniques to solve real-world problems. ------- Я выпускник City University of Hong Kong, специализирующийся на машинном обучении, больших языковых моделях (LLM) и обучении с подкреплением. Имею опыт работы исследователем и AI Solution Engineer, где занимался разработкой систем на основе LLM, RAG и масштабируемых ИИ решений. Увлечён исследованиями в области искусственного интеллекта, разработкой эффективных моделей глубокого обучения и применением передовых технологий для решения реальных задач.

Career

Experience: 4 years

Experience

AI Solution Engineer

Witernet Ltd. Hong Kong

Nov 2024 — Jun 2025

• Built a FAISS-based vector search and RAG pipeline, improving answer accuracy by 22% on internal evaluations. • Implemented a word-embedding recommendation system • Deployed a 32B-parameter LLM with vLLM, building an inference API and optimizing latency

Research Assistant

City University of Hong Kong

Jun 2025 — Present

• Co-authored and presented an IEEE TALE 2025 paper on a new RL-based LLM training method for Socratic tutoring in introductory programming education. • Fine-tuned 7B-parameter language models using PPO, DPO, and GRPO; integrated RAG for retrieval-augmented tutoring and distributed training on an 8-GPU cluster with DeepSpeed ZeRO-3. • Built a custom training dataset by combining public sources with LLM-generated examples for Socratic tutoring and student-assessment tasks. • Contributed to the design of CityU’s Introduction to Computer Programming course.

Research Assistant

Institute of Smart Systems and Artificial Intelligence, Nazarbayev University

Jun 2023 — Jul 2023

Built a ResNet-50-based CNN image classifier and boosted recognition accuracy by 12% using regularization and AI-driven data augmentation.

Education

BSc Computer Science

City University Of Hong Kong

2024 — 2028

Had an Exchange Semester at the University of Birmingham

Selected projects

Diffusion Models for Image and Text Generation

• Implemented a latent diffusion training/inference stack in pure PyTorch: VAE encoder/decoder, text encoder (CLIP-based), DDPM sampler, and end-to-end data + training pipeline • Extended the framework to text generation by implementing a diffusion language model, applying the denoising objective to discrete token sequences

Open project

MoE Language Model from Scratch

• Designed and trained a custom MoE language model from scratch in pure PyTorch, scaling distributed training across 6 GPUs with DDP. • Fine-tuned for reasoning with GRPO (also, pure PyTorch) • Matched performance of best-reported sub-150M parameter models, scoring 27.2 on ARC-C and 52.8 on Winogrande (vs. 27.1 and 53.1)

Open project

Checkers Engine using Reinforcement Learning

• Trained a DQN bot via self-play to learn checkers from scratch in pure PyTorch, reaching 99% win rate vs. a heuristic opponent.

Open project

Skills

PythonPyTorchMachine LearningDeep LearningLLMsReinforcement LearningGenerative AIRAGDistributed TrainingHugging FaceDeepSpeedvLLMFAISSLoRATensorFlowC++SQL