@sultan-ars
Junior Machine Learning Engineer | Applied AI
I like building AI tools that have a clear job, like checking cashier footage for possible violations or pulling quiz schedules from 15 different sites. I’m studying Applied AI at Astana IT University.
Junior Machine Learning Engineer | Applied AI
Built a service that analyzes bank cashier CCTV footage and flags possible cash-handling violations with timecodes, 1 to 3 evidence frames, and National Bank of Kazakhstan regulation references. A two-stage Qwen3-VL-30B-A3B pipeline analyzed 10-second sliding windows at 2 fps across 13 violation types, then rechecked error-prone types on zoomed crops. Benchmarked five vision-language and embedd...
Loaded and modeled 100K+ real orders from the Olist dataset in PostgreSQL across 8 related tables. Performed exploratory data analysis with pandas and matplotlib, wrote analytical SQL using joins, window functions, and aggregations, and built an interactive Apache Superset dashboard with charts for revenue trends, top categories, ratings, and payment methods.
Built a Streamlit chat app for natural-language questions across uploaded PDFs. Implemented a LangChain RAG pipeline with PyPDF2 extraction, chunking, OpenAI embeddings, FAISS semantic retrieval, and conversation history for multi-turn Q&A. Containerized the app with Docker for local and portable deployment.
Built an async Telegram bot that merges quiz schedules from 15 differently structured websites into one database and serves them through /today and /week. It refreshes stale data on demand, prevents duplicate runs with an asyncio lock, caps concurrent Playwright parsers, handles date edge cases including Almaty UTC+5, and includes admin tools for manual entry and a daily channel poll.
Nazarbayev University
Nazarbayev University is a research university in Astana, Kazakhstan, established in 2010, with programs in science, engineering, and graduate research.
May 2025 — Feb 2026Built a weakly supervised wheat disease forecasting pipeline, processing 666,000+ hourly records and comparing XGBoost, LSTM, and GRU across climatic regions.
• Collected and cleaned data for a weakly supervised time-series pipeline for wheat disease forecasting, processing 666,000+ hourly records across multiple locations. • Engineered rolling weather features using 3 to 72-hour windows; trained and tuned XGBoost, LSTM, and GRU models to compare forecasting approaches and evaluate generalization across climatic regions.
Nazarbayev University
2022 — 2026Astana IT University
2026