ML Engineer interested in applied AI roles
Computer Science student passionate about ML/AI, with hands-on experience in Computer Vision, LLM optimization, and applied research, focused on building efficient real-world AI solutions
Nazarbayev University
A research university in Kazakhstan.
Mar 2025 — Aug 2025Fine-tuned and evaluated language models, reporting 90%+ performance and analyzing accuracy-energy trade-offs across 12+ combinations.
• Fine-tuned TinyLLaMa/BERT/ALBERT on SNIPS/RTE using FT/LoRA/BitFit, reaching 90%+. • Evaluated 12+ model-technique combinations and analyzed accuracy-energy trade-offs with metric SAM.
KAI DEEP TECH
Aug 2025 — Sep 2025Built a multilingual n8n agent workflow and a pluggable YOLOv8 implementation, reaching 0.95 mAP@0.5 on a custom dataset.
• Built a minimal multi-agent workflow in n8n routing KK/RU/EN queries and orchestrating LLM calls. • Implemented a YOLOv8 from scratch with pluggable backbones/necks for fast domain adaptation and accuracy gains. • Achieved 0.95 mAP@0.5 and 0.57 mAP@0.5:0.95 on a custom 4-class dataset.
iSwim
Nov 2025 — PresentMaintains a computer-vision pipeline across four pool-camera streams, reducing false alerts by 25% and keeping alert latency under two seconds.
• Maintained a production pipeline using YOLO, OpenCV, and GStreamer across 4 pool-camera streams. • Tuned detection, tracking, and drowning-risk thresholds, reducing false alerts by 25% while preserving performance. • Optimized frame-handling components to sustain 20+ FPS and keep safety-alert latency below 2 seconds. • Supported CV integrations with Redis, FastAPI, and Docker, ensuring reliable video streaming and alert delivery.
Nazarbayev University
2023 — 2027Built a face-registration and check-in desktop app with a computer-vision pipeline, SQLite persistence, liveness challenges, face embeddings, and a headless CV core.
Open projectFine-tuned Whisper with LoRA on the UASpeech dysarthric speech corpus, reducing WER from 142% to 38.5%; a vocabulary-constrained normalizer brought final WER to 33.7%. Built an ASR-to-TTS restoration pipeline.
Open projectCollected and cleaned content from 20+ Central Asian recipe websites, built semantic search and Gemini-powered RAG answers, and deployed a Telegram bot that reduced empty responses by 40%.
Open projectPrepared a 5,000-movie dataset, generated Llama2 text embeddings, built a FAISS vector database, and launched a Streamlit recommendation app with 70% user satisfaction.
Open project