N.Omerbayev // ds·llm·agents
N.Omerbayev // ds·llm·agents
5+ years in data science. Currently building agentic RAG and text-to-SQL systems on self-hosted open-weight models, alongside deep-learning demand forecasting and real-time computer vision on edge hardware.
Production ML systems across generative AI, forecasting and computer vision — from self-hosted LLM agents to edge inference on custom hardware.
Agent systems that give business users natural-language access to enterprise knowledge and data. A retrieval-augmented pipeline over internal documentation and warehouse metadata, plus a text-to-SQL agent that turns business questions into validated SQL against the analytical warehouse — cutting reliance on manual ad-hoc reporting. Open-weight models (Qwen-class, ~27B) are served fully on-premise on in-house GPU infrastructure, with quantization and inference optimization. Agentic workflows use function calling to query data sources and chain multi-step analytical tasks.
A demand-forecasting model on a Temporal Fusion Transformer (PyTorch Forecasting + Lightning) with a 56-day horizon and an 84–112 day encoder. Quantile outputs (P2–P98) give uncertainty estimates, and a custom horizon-weighted quantile loss prioritizes near-term accuracy. Features cover promotions, log-scale pricing, weather, holidays, store format and city, with a dedicated model branch for new SKUs that have no sales history. Trained at scale with per-epoch sharded streaming data loading and mixed-precision (bf16) on GPU.
A real-time computer-vision pipeline running on an edge device: camera → YOLO inference on a Hailo accelerator → object tracking → zone and line counting, persisted to a database. The BYTETracker implementation is optimized with Numba (JIT IoU, cached Kalman state) for high-throughput multi-object tracking. Robust frame handling covers letterbox preprocessing, NMS, class stabilization, dwell-time and IN/OUT line-crossing logic, with GStreamer camera ingestion and automatic restart.
License-plate detection and character reading with PaddleOCR, integrated into the edge video-analytics pipeline so detection, tracking and plate reading happen in a single on-device stream.
A churn-prediction system that identifies at-risk customers in retail / e-commerce, driving reactivation of 8% of the inactive client base through targeted outreach. Part of broader customer-analytics work spanning uplift modeling and demand forecasting.
Five years across retail, banking and telecom — building ML that reaches production. Roles held as Data Scientist throughout.
Building agentic RAG and text-to-SQL over enterprise data on self-hosted open-weight LLMs, plus TFT-based demand forecasting for a large grocery retail network.
Churn prediction (8% reactivation), demand forecasting for 15,000+ SKUs at 70%+ accuracy, uplift modeling (+3% conversion), and automated SQL-based reporting.
Real-time fraud detection on streaming data via a Kafka broker, and a B2B sales-boost prediction matrix enriched with parsed open data (tax & national statistics).
Identified new clients and revenue streams by parsing geocoordinates of 20,000+ legal entities from open sources, visualized on interactive HTML maps.
I regularly take part in applied ML competitions — a way to stay sharp outside the day job and benchmark my solutions against strong baselines on unfamiliar data. Most recent below.
Deciding whether two marketplace listings describe the same physical product — matching across noisy titles, inconsistent attributes and product images at catalogue scale. Work centred on text and image embeddings, candidate retrieval over a large item space, and a ranking model on top of the candidate pairs.
A recommender-system challenge on large-scale VK interaction data: predicting user engagement from sparse implicit feedback. Focus on candidate generation, behavioural and temporal feature engineering, and gradient-boosted ranking over the retrieved candidates.
RAG, text-to-SQL, function calling / tool use, agentic workflows, LangChain, self-hosted open-weight LLMs (Qwen), quantization, GPU serving, vector search & embeddings
Temporal Fusion Transformer, PyTorch Lightning, time series, quantile forecasting
YOLO, PaddleOCR (ANPR), BYTETracker, Hailo edge inference, OpenCV, GStreamer
Classification, regression, clustering, churn, uplift modeling, NLP, A/B testing
Python (Pandas, NumPy, scikit-learn, PyTorch), SQL, PySpark, Spark, Hadoop, Hive, Apache Kafka, Airflow
AWS SageMaker / EC2 / Presto, Docker, Git, Linux, MariaDB / PostgreSQL, Raspberry Pi / edge