Mukhambet Tolegenov

AI Developer

Neurotech

Built search, RAG, intent classification, and media monitoring for a 91,000-document state archive, with measured gains in retrieval and classification quality.

• Built hybrid search using BM25 (Elasticsearch), bge-m3 (Qdrant), reciprocal rank fusion, and a blend with Qwen3-Reranker-0.6B on vLLM. nDCG@10 rose from 0.465 with BM25 alone to 0.734 on a set of 50 queries and 5,500 judgments. • Built a LangGraph RAG assistant with streaming through FastAPI WebSocket, a short history of three turns, Kazakh-to-Russian translation, issai/Qolda responses with links, and fallback modes for failures.

Research Assistant

Anglia Ruskin University

A public university in the UK, with campuses in Cambridge and Chelmsford and a broad range of teaching and research programs.

Improved audio emotion recognition accuracy from 38% to 59% and developed CNN models for facial emotion recognition; the work appeared in Springer LNCS.

• Improved an audio emotion recognition model's accuracy from 38% to 59% using an attention mechanism. Attention also improved the audio-visual model by 7%. The work was published in Springer LNCS. • Developed CNN models for facial images in an emotion recognition pipeline, achieving 55–63% accuracy.

Python Developer

KMG Engineering

Automated data transfer from more than 20 fields to a central database with a Python microservice.

• Automated data transfer from more than 20 fields to a central database using a Python microservice.

IT Specialist

Chevron

A global energy company active in oil, gas, and other energy businesses, with operations and customers around the world.

Automated two to three processes in Power Automate, saving about five hours of manual work per week.

• Automated two to three business processes in Power Automate, reducing manual work by about five hours per week.

Master's degree, Artificial Intelligence

Anglia Ruskin University

Bachelor's degree, Big Data Analytics

Astana IT University