• Built RAG based Generative AI Assistants to assist client-side users via LLMs. Used Streamlit in order to provision a user-friendly interface which users can select a subject and ask their questions instead of reading all the related documents to find answers. This reduced HR's workload by 15%. • Built an ML Interface application that provisions users hands-on classification, regression and time series model training and visualizations using Scikit-Learn and Streamlit. The training processes are logged, visualized and combined as a report in an HTML format. • Wrote scripts to increase efficiency in LLM prompts. Removed unnecessary tokens before processing, with this approach LLMs token cost is reduced by 30%, this led to more cost-efficient processing. • Trained a time series model via utilizing fully managed AWS SageMaker and S3 services, serialized the trained model, created dynamic ETL pipelines and deployed the target designed DAG on destined production-ready Apache Airflow platform.

• Used Blazor to build interactive web pages and ASP.NET Core for server-side logic.• Stored user data in a SQL database and worked with basic database queries.

• Analyzed one-year period of tabular data on Tableau via applying exploratory data analysis techniques and utilizing statistical approaches to identify anomalies.• Utilized FacebookProphet in order to forecast future product sales and visualized the statistical results using Seaborn and Plotly. To focus on the model's sensitivity, used R2 and RMSE as the performance indicator.

• Contributed to the documentation of a large international open-source machine learning framework project by testing functions and optimizing performance using Python and Git.

• Built Object Detection Models to detect surface damages in solar panels and wind turbines. Used PyTorh and TensorFlow as the main code, and used AlexNet as the CNN architecture. For the training and deployment process, utilized AWS SageMaker as the PaaS.• Analyzed air pollution data around factories that is acquired from the remote controlled drones and visualized the hazardous air on the map. Filed the analysis as a report and proposed solutions to the factory on how we can reduce pollution.

• Built line detection algorithms using Hough Lines technique and improved the detection by applying image processing methods such as dimension reduction with greyscaling and image sharpening. • Trained Object Detection Models to detect traffic signs and traffic lights in various lighting circumstances using DarkNet YOLOv4 with an average accuracy of ~90\%. The trick behind this accomplishment is the data was prepared 100\% manually. All the images that are used to train the model were taken with different angles in different day hours with specifically chosen traffic sign and traffic light designs.• Both improved line detection algorithm and trained Object Detection Model is implemented into the Nvidia Jetson Nano using the Jetson Inference modules. With this newly built system, Cukuruva Hydromobile Team landed in \href{https://bilimgenc.tubitak.gov.tr/makale/efficiency-challenge-elektrikli-arac-yarislarinda-final-heyecani}{\underline{2nd placement in the 2020 TÜBİTAK Efficiency Challenge.}}

• Collaborated on building a Deep Learning User Interface that consists tasks of Regression, Classification and Time Series Forecasting. Model architectures heavily depends on Tensorflow and Scikit-Learn. Interface's available architectures: RNN, MLP, SVM, RandomForest, LGBM, CatBoos, XGB, Ridge, Linear Model, GRNN, SARIMA, ELM, RandomWalk, Moving Average.• Contributed by collecting specific research papers and patents, and summarizing their concepts and use-cases. Utilized collected research papers and patents to create a new approaches and projects.

• Using applicants' information data, built a XGBoost classifier to classify candidates who may submit their applications. In order to deploy the model, provisioned a Serverless AWS SageMaker Inference. Thus, HR's candidate tracking cost reduced by 30%.