Trainee, NLP Specialization
EPAM Systems
Global IT services and software engineering company.
Mar 2026 — PresentBuilt document NLP, information extraction, semantic retrieval and RAG systems; reported 97% entity-level F1 and improved retrieval metrics.
• Built an end-to-end OCR and document understanding pipeline using Tesseract, Docker, and PAWLs, converting PDFs into training-ready datasets for downstream NLP tasks.
• Developed NER and information extraction systems for invoices using BERT, LayoutLMv3, and LLM-based extraction, achieving entity-level F1 of 97%.
• Fine-tuned BERT for text classification and document understanding using HF and PyTorch, achieving 12% improvement in Macro F1 over the baseline model XGBoost.
Research Assistant, Computer Vision
Nazarbayev University
Research university in Kazakhstan.
Sep 2025 — PresentDeveloped and served a skin lesion segmentation model; the work was accepted at IEEE WCCI 2026 (FUZZ-IEEE).
• Developed a skin lesion segmentation framework combining ResNet34 and a Fuzzy Self-Attention decoder, reaching SOTA IoU and Dice scores.
• Deployed the model through a FastAPI-based inference service and exposed it with Ngrok to support interactive demos and remote testing.
• Published this work as FSSM-ResNet: Hair-Aware Skin Lesion Segmentation with Embedded Fuzzy Structural Suppression, accepted at IEEE WCCI 2026 (FUZZ-IEEE).
Computer Vision Intern
Sergek Group
Kazakhstan-based smart city technology company, including traffic monitoring and intelligent transportation systems.
Apr 2025 — Sep 2025Optimized monocular depth estimation for edge devices, halving reported error metrics and achieving up to 4x faster inference.
• Developed real-time monocular depth estimation models for smart city systems, improving AbsRel from 0.18 to 0.09 and RMSE from 6.2 to 3.1.
• Distilled Metric3D models to lightweight MobileNet backbones and optimized deployment with ONNX and TensorRT, achieving up to 4x inference speedup (120 ms to 30 ms/frame) without accuracy loss.
• Deployed optimized models on NVIDIA Jetson devices for edge inference in real-time smart city applications.
• Explored and evaluated Vision-Language Models (VLMs) for combining visual understanding with textual reasoning in applied computer vision workflows.
Research Assistant, Speech and Language
ISSAI
Institute of Smart Systems and Artificial Intelligence at Nazarbayev University.
Jul 2023 — Oct 2024Worked on Kazakh ASR and voice conversion; reduced ASR WER and latency and contributed to a 557-speaker dataset.
• Fine-tuned Meta’s Seamless Streaming model for real-time Kazakh ASR, reducing WER from 23.5% to 19.6%.
• Designed a streaming-aware training and inference pipeline, reducing end-to-end latency from 480 ms to 230 ms and improving real-time factor from 0.92 to 0.75.
• Developed Kazakh neural voice conversion systems using Transformer-based models, HiFi-GAN, and FreeVC.
• Built a PyTorch-based speech pipeline and contributed to a 557-speaker Kazakh dataset, supporting downstream speech and language modeling experiments.
