Skip to content

Auyeskhan Alibekov

@engineer

AI Engineer - NLP, LLM fine-tuning, RAG, vLLM

Data Scientist with 3 years of experience in NLP for Kazakh and Russian. I fine-tune LLMs, build multilingual RAG pipelines, and deploy models on-prem with vLLM.

Open to

Finding a mentorFunding & grantsStartup collaborationProject collaborationLearning with others

About

First-authored paper on a multilingual AI avatar coming out in September. Open to ML/AI Engineer roles in Kazakhstan or remote.

Career

Experience: 2 years

Experience

AI Engineer (Contract Project)

[EMPLOYER]

Aug 2026 — Present

Pilots GBrain agent memory for BRK, deploying a local stack and progressing a migration to PostgreSQL with pgvector.

• Evaluating and piloting an open-source agent memory layer (GBrain) for an enterprise client: local deployment on Bun and Ollama with bge-m3 embeddings (1024-dim) over a PGlite store, with migration to PostgreSQL and pgvector in progress. • Driving adoption of an unfamiliar stack from scratch, using Claude Code and MCP tooling as the development environment; working directly with the client’s team on a daily cadence.

Summer Research Program — Mentor

ISSAI (Institute for Smart Systems and Artificial Intelligence) – Nazarbayev University

Jun 2026 — Present

Mentors a team integrating LLM, ASR, and TTS for a multilingual real-time AI avatar demo.

• Mentoring participants building a multilingual AI avatar: integrating LLM, ASR, and TTS into one real-time pipeline and assembling an end-to-end demo.

Data Scientist

ISSAI (Institute for Smart Systems and Artificial Intelligence) – Nazarbayev University

Jun 2024 — Present

Builds and evaluates Kazakh-language AI systems, including LLM serving, safety testing, RAG, ASR fine-tuning, and dataset work; reports 87% QA accuracy, 15% lower WER, and 23% fewer hallucinations.

• Deployed and operated self-hosted LLM inference servers using vLLM (Qwen, InternVL, 3B–30B scale), exposing API endpoints on an HPC/SLURM GPU cluster for lab-wide benchmark and evaluation workloads. • Designed an automated safety evaluation harness for Kazakh-language LLMs (KazSHIELD), running static and dynamic test suites against vLLM endpoints on every checkpoint. • Built multilingual RAG pipeline using Haystack and FAISS with hybrid retrieval, reaching 87% QA accuracy on Kazakh/Russian benchmarks.

Research Assistant

ISSAI (Institute for Smart Systems and Artificial Intelligence) – Nazarbayev University

Feb 2024 — Jun 2024

Developed OCR pipelines for Kazakh legal PDFs and integrated embedding retrieval into RAG systems on a GPU cluster.

• Developed and tested OCR pipelines using Docling for Kazakh legal PDFs with structured Markdown output. • Integrated embedding retrievers into RAG systems; automated server-side inference pipelines on GPU cluster.

Education

B.S., Software Engineering

Astana IT University

2020 — 2024

Selected projects

Multilingual Generative AI Avatar

Led development of a real-time multilingual avatar system in Kazakh, Russian, and English, with local and cloud deployment modes. Integrated LLM inference, ASR, TTS, and facial animation into a streaming end-to-end pipeline that ran live in demos.

Open project

Qolda-Math

Co-developed a small vision-language model for mathematical reasoning in Kazakh, evaluated on GSM8K and MathVista. Contributed to dataset construction, the HPC training pipeline, and comparative benchmark evaluation against baseline Qolda models.

Open project

Skills

PythonSQLBashLLM serving (vLLM, 3B–30B)LLM fine-tuning (PEFT/LoRA)TransformersRAGModel evaluation and benchmarkingASR/TTSOCREmbeddingsLangChainHaystackQdrantFAISSClaude CodeMCPAgent Memory (GBrain)Ollamabge-m3 Embeddings