Leadership & Development: • Developed the teye-va library (C++, Lua) in a team of 3 engineers, a core video analytics (VA) module of the main product• Contributed to full cycle ML models development (business understanding, data collection, quality, model training, inference optimization and further integration)• Designed MLOps pipeline for automatic model retraining to decrease time to market substantially• Designed and implemented ML inference pipelines and APIs for 4 modules: face recognition, license plate recognition, unattended object detection, and situational analytics. Conducted unit and end-to-end testing to ensure system robustnessGenerative AI: • Researched and contributed to finetuning Pali-Gemma (Vision-Language Model) for large-scale dangerous situation detection via Visual Question-Answering• Surveyed recent LLMs (e.g. Llama, Mixtral), RAG-based techniques, and parameter-efficient finetuning strategies for improving user experience with the VA systemResearch to Production: • Reduced false positive rate of the YOLO model by 80% through the development of an uncertainty estimation method and addressing data-related issues• Improved CRNN model’s accuracy from 90.2% to 98.3% by utilizing data-centric and architectural approaches• Achieved 20× reduction in latency with TopFormer (CNN-Transformer) (2) increased illumination robustness by designing inter-frame difference augmentations

• Developed a novel uncertainty estimation method for image segmentation that achieves 55% and 23% improvement in correlation with expert disagreements at the image and pixel levels

• Developed a novel method for improving the generalizability of CNNs to stain changes (pathology) in semantic segmentation with an improvement of 19% in the Dice score on average. Validated results with 5 external test sets, 2 encoders, and interpretability analysis• Improved prediction performance of a tree-based ensemble model for air pollutant (PM2.5) estimation (51% increase in R2, 53% down in RMSE) by preprocessing and fusing a raw set of NASA satellite images and NCEP climatic records, and additional feature engineering. Identified action policies for decreasing the pollutant levels concerning the model predictions• Investigated feasibility of predicting genetic marker mutation from MRI scans by implementing custom CNN for fusing feature representations of two modalities (+0.12 AUC), and adapting existing network architectures• Contributed to preparing papers (3 accepted) for submitting aforementioned results in different teams

• Developed data cleaning and merging scripts for heterogeneous agricultural data from 14 main regions in the country. Identified how efficiently agribusinesses were using allocated government subsidies by clearly visualizing findings with Power BI• Developed multi-factor simulation for sugar price fluctuations at region level by performing correlational analysis, extracting feature importances from Random Forest, and implementing inverse MLP model. A ministry representative approved the solution for further integration• Adapted DELPHI model from MIT for local covid-19 predictions at city and town levels by modifying government response indicators and investigating relevant research papers• Participated in preparing presentations to government officials in different teams for each project

• Developed dynamic scripts for merging, database backup, and exception handling for OLAP cube

• Selected among 118 candidates from 2000 applicants country-wide. Learned practical aspects of ML, Project Management, and Front-end development during the first month• Implemented UI for Alpha Way with Vue, a web application for tracking employees’ progress (profile, tasks, skills, examination, and career development options). A bank acquired our team’s solution

• Designed and constructed different websites for local companies using HTML, CSS, JavaScript, Bootstrap, PHP, MySQL• Helped businesses with advertising, and promotion using Google Ads