Ravelos

• Developed intermittent demand forecasting systems for retail-like datasets containing 20K+ products, 3M+ rows, and approximately 85% zero-demand observations.
• Designed a two-stage forecasting architecture using classification and LightGBM-based regression pipelines for sparse demand modeling.
• Built fully recursive inference workflows where future predictions were generated without access to unavailable future information, preventing data leakage.
• Improved forecasting accuracy from approximately 40% to 65-70% using custom business-oriented evaluation metrics and advanced feature engineering.
• Reduced end-to-end ML pipeline runtime from 40 minutes to 4 minutes through vectorization, parallelization, memory optimization, and matrix-based computation.

Vcamp

• Served as team captain for the TEKNOFEST Aviation AI Competition, defining the overall architecture and directing parallel workstreams across object detection, position estimation, and reference-object matching.
• Designed and implemented a hybrid visual odometry pipeline combining ORB and optical flow with RAFT-based motion estimation, confidence-aware arbitration, and an iterative extended Kalman filter, achieving approximately 7.5 FPS on custom drone footage.
• Built an end-to-end reference-object matching pipeline for aerial RGB and thermal imagery, fine-tuning YOLOv8 across seven training rounds and reaching 0.749 mAP50 at approximately 36 ms per frame.
• Directed a four-class UAV object-detection subsystem and defined its dataset architecture, training strategy, and validation criteria for a dataset containing more than 35K images.
• Architected a unified real-time competition client that integrated all three subsystems behind a modular inference pipeline and produced JSON submissions compatible with the competition API.
• Built iteration tooling including a browser-based annotation editor and a temporal label-densification technique for scaling small hand-labeled datasets.
• Co-developed the Vcamp website in a five-person engineering team, building Python data models and REST APIs and integrating them with a React and Next.js frontend.
• Implemented multilingual content, pagination, certificate validation, a dynamic team gallery, responsive UI improvements, and fixes for more than 30 tracked issues.