Senior LLM Engineer
WebDesignSun
Dec 2025 — PresentShipped production vLLM serving, custom guided decoding, evaluation pipelines, and canary rollouts with cloud GPU operations.
• Built and shipped a production LLM inference service on vLLM, fronted by a chunked, OpenAI-compatible API proxy and rolled out via staged canary A/B on live traffic. vLLM, FastAPI, Docker.
• Implemented custom guided decoding as a batched, lockstep-safe vLLM logits processor (auxiliary model with per-step KV-cache management), working deep in vLLM and Hugging Face Transformers internals. vLLM, Transformers.
AI Engineer
Finkey
Jul 2025 — Apr 2026Built a feedback-based RAG pipeline and structured NL-to-SQL system, and architected its production agent and API infrastructure.
• Built a retrieval-augmented generation (RAG) pipeline over a curated store of user-query/SQL pairs scored +1/−1 by feedback: at inference the system retrieved the most similar past queries and injected them as few-shot context, replacing a static prompt-only design. RAG, semantic retrieval.
AI Research Scientist
Astana IT University
A university in Kazakhstan focused on information technology and related fields.
Jan 2025 — PresentFine-tunes and benchmarks open-source LLMs and vision-language models against CNN baselines.
• Fine-tuned open-source LLMs and vision-language models (Qwen2, Qwen2.5, Llama-3.2, SmolVLM) with LoRA, and benchmarked them against CNN baselines across data regimes. PyTorch, Transformers, LoRA.
Computer Vision Engineer
ReLive Research
Oct 2019 — Dec 2025Built and deployed real-time ML systems, optimizing inference to 8 ms with TensorRT.
• Delivered real-time ML systems end-to-end — trained models, optimized inference to 8 ms with TensorRT, containerized with Docker, and served via FastAPI in production. TensorRT, Docker, FastAPI.
Research Assistant
Nazarbayev University
An autonomous research university in Kazakhstan.
Sep 2021 — Dec 2024Conducted research in self-supervised image segmentation and risk-averse optimal control; co-authored a conference paper.
• Co-authored self-supervised representation learning (BYOL) for image segmentation, leading to a conference publication. PyTorch.
• Proposed and implemented an Approximate Dynamic Programming model for risk-averse optimal control (MSc thesis) — sequential decision-making under uncertainty. Python, optimization.
