Amir Izimov

AI / ML Engineer

Yurt Tech

Built and optimised production video anomaly detection and multi-camera inference; achieved 862 FPS and reduced model latency and VRAM use.

• Built and optimised a production CCTV anomaly-detection service on an 8B video-LLM (InternVideo3, SGLang, fp8 on H200): designed a multi-stage classifier with logprob-based verification gates that suppressed false positives to ∼100% precision, and cut per-clip inference latency ∼40% by profiling and fixing the model’s prefill dispatch path. • Own the real-time inference stack for multi-camera analytics: native ByteTrack multi-object tracking, cross-video pose estimation batching (up to 128 crops) with TensorRT serving, and fused SCRFD preprocessing for concurrent CCTV workloads.

AI / ML Engineer

BelesAI

Led CV architecture for a multi-stream production platform and improved cross-camera person Re-ID accuracy from 30% to 90%.

• Lead engineer on a production real-time CV platform over multiple concurrent RTSP streams: detection, face recognition, attribute classification, and per-camera zone & dwell-time tracking; Dockerised multi-GPU deployment. Defined the CV architecture end-to-end. • Designed a cross-camera person Re-ID and tracking pipeline (YOLOv8 + OSNet + FAISS, 8 concurrent RTSP streams, accuracy 30%→90%) with an RTMPose pose-based behaviour-detection cascade and automated incident clip export.

ML / AI Engineer (Contract)

Big Dream Lab

Delivered a production RAG document-QA system for KEGOC, reporting 90% accuracy and 60% lower latency.

• Built and deployed a production RAG document-QA system for KEGOC (national electricity grid operator): 300+ documents, pgvector embeddings, BGE reranking, LLM generation on H200 GPUs; 90% accuracy, 60% latency reduction.

Computer Vision Intern

SERGEK Group

Trained and deployed YOLO detectors for six vehicle categories on a 1.3M+ image dataset, reaching 90% production accuracy.

• Trained, benchmarked, and optimised YOLO detectors on a 1.3M+ image dataset at a large-scale road-monitoring operator: 90% production accuracy across 6 vehicle categories. Dataset curation, model evaluation against baselines, hyperparameter search with Optuna, and production deployment.

Research Assistant

HF Measurement & Characterization Laboratory, Nazarbayev University

Contributed to research on GWO-optimised XGBoost for GaN HEMT prediction, reporting 99.9% accuracy.

• Co-authored a research paper on GWO-optimised XGBoost for GaN HEMT device prediction (99.9% accuracy): literature review, rigorous benchmarking against baselines, and result interpretation.

B.Eng., Electrical & Computer Engineering

Nazarbayev University