Yerdauit Torekhan

Lead Computer Vision Researcher

Qoima Automation

Designed and deployed a computer vision system from scratch, cutting inference latency 5–6x and supporting 15 camera streams on one machine.

• Joined as the project’s sole ML developer; designed the computer vision architecture (detection, multi-object tracking, vision-language models) from scratch and implemented it end to end. • Optimized inference with TensorRT, FP16/INT8 quantization, and NVIDIA Triton Inference Server, cutting latency roughly 5–6x; the full system decodes 15 camera streams with FFmpeg and runs on a single machine with 16 GB of VRAM.

Middle Computer Vision Researcher, Face Anti-Spoofing

OZ Forensics

Developed and benchmarked lightweight face anti-spoofing models for a platform handling 100M+ verifications monthly, with evaluation across attacks and demographic slices.

• Worked on face presentation-attack detection (PAD) for an identity-verification platform handling 100M+ verifications per month; designed and trained an original lightweight LSTM-based model, adapting an approach from a recent paper. • Benchmarked candidate models with ONNX Runtime and TensorRT (FP16/INT8); the model trains on CPU and was the fastest among the team’s candidates while keeping competitive accuracy. • Ran training and evaluation on datasets of millions of videos on GCP, tracking experiments in Weights & Biases; generated affine-transformed video augmentations with FFmpeg.

Junior Computer Vision Engineer

BTS Digital

Created the team’s first synthetic data pipeline and worked on face anti-spoofing models and a domain-specific RAG system.

• Built the team’s first synthetic training-data generation pipeline with ComfyUI, running entirely on local infrastructure; before it, no data-generation tooling existed. • Developed and evaluated face presentation-attack detection models for real-world spoofing scenarios, tracking APCER, BPCER, and ACER under ISO/IEC 30107-3. • Built a domain-specific RAG system using LangChain, custom tokenization, vector search, and a DeepSeek-based generator.

Data Engineer

Adata

Built Talend and Python ETL pipelines, refactored processing bottlenecks, and added data cleansing and validation.

• Developed ETL pipelines with Talend and Python; refactored Java/Python processing bottlenecks and implemented pandas-based data cleansing and validation.

Research Engineer

FLIS Lab, Kazakh-British Technical University

Led visual feature work for a first-author IEEE SIST 2025 paper and contributed to emotion recognition and fuzzy color research.

• Led visual-feature extraction and evaluation for a first-author study on painting aesthetics, published at IEEE SIST 2025. • Contributed to research on multimodal emotion recognition and fuzzy color representations, including model implementation, experiments, and manuscript preparation. • Mentored undergraduate researchers in experimental methodology and reproducibility.

M.Sc., Data Science

Kazakh-British Technical University (KBTU)

B.Sc., Computer Information Systems

Kazakh-British Technical University (KBTU)