Skip to content
DSML.KZ
ZK

Zhaidar Kairat

@EredirrosConnected with DSML. This person's Telegram account has been confirmed through DSML Kazakhstan.

Machine Learning and Computer Vision Engineer

ML engineer and AI researcher specializing in computer vision and medical imaging. Interested in deep learning, medical image analysis, and developing intelligent healthcare applications.

Open to

Joining pet projectsJoining competitions & hackathonsPaper discussionsProject collaborationResearch collaboration

Career

Experience: 2 years

Experience

ML / Computer Vision Engineer (Research)

Astana IT University

A university in Kazakhstan focused on information technology and related fields.

Jul 2025 — Present

Built camera vision prototypes and medical imaging models, then productionized GPU inference with monitoring and expanded CI tests from 16 to 84.

• Built two camera-based CV prototypes (flour-bag counting on a conveyor line, parking occupancy from overhead cameras): data collection, annotation, training and evaluation on real footage. • Trained CT and MRI segmentation models and served them as an async GPU inference service (FastAPI, Celery, RabbitMQ, Redis, Docker). • Hardened it for production: preflight and readiness probes, idempotent jobs, 429/503 admission control, typed errors, tests 16 to 84 in CI, Prometheus/Grafana monitoring and a Telegram alert bot.

Data Scientist Intern

Kazakhstan Housing Company

A Kazakhstan housing finance and development institution.

Dec 2024 — May 2025

Built cash-flow risk simulations and housing market forecasts, and raised credit-scoring accuracy from 0.81 to 0.94.

• Built Monte Carlo simulations for cash-flow risk evaluation and time-series forecasts of the housing market; raised credit-scoring accuracy 0.81 to 0.94 (gradient boosting, SMOTE).

Education

M.Sc., Applied Artificial Intelligence

Astana IT University

2025

B.Sc., Big Data Analysis

Astana IT University

2022 — 2025

Selected projects

Flour Bag Counting from Video

Industrial conveyor line with live RTSP cameras. YOLOv8x detection with ByteTrack counts each bag as its track crosses a virtual line, with an end-of-track pass to recover missed bags. Trained on 34,660 frames after cleaning 157,905 raw annotations. Smaller YOLOv8 models were deployed with OpenVINO INT8 on 8 CPU cores. YOLOv8n INT8 reached 4.9 ms per frame at mAP50 0.913. A live pipeline for 1-...

CT Lymph Node Detection and Breast MRI Tumour Segmentation

For CT, designed a 2.5D Cross-Slice Transformer with 7.97M parameters, the best of the 2.5D models under one fixed protocol and level with 3D nnU-Net on small nodes, using 296 fully annotated scans. For MRI, benchmarked six architectures on 1,500+ multi-phase breast MRI studies and found that kinetic signal, not spatial structure, was the dominant discriminator.

Home Video Search

Zero-shot text-to-video search over 1,863 Charades home videos using SigLIP2 frame embeddings and Qwen3-VL-30B clip descriptions in pgvector. VLM reranking raised Recall@1 from 0.30 to 0.56 on 1,000 test queries. An LLM agent uses search, localisation and clip-inspection functions to return a clip and time range for a text query.

Kazakhstan Banking Intelligence

On-premise AI risk assistant for Kazakh banking data. A tool-calling agent using Qwen3.8-27B on vLLM searches 6,300 documents and 530k figures with hybrid retrieval and reranking. Accuracy on 50 expert questions rose from 40% to 98%, with 32/32 on a held-out set.

Skills

PythonC++SQLPyTorchOpenCVUltralytics YOLOByteTrackMONAInnU-NetSigLIP2 image embeddingsEmbedding similarity search (pgvector)Scikit-LearnOpenVINOONNX / ONNX RuntimeINT8 quantization (NNCF)ffmpegRTSPH.264LinuxGit