Dogugun Ozkaya

Served as the lead AI engineer and built the AI infrastructure for the company, including the local Qwen3 model, OpenWebUI, and Elastic. Introduced agentic RAG methodology for QA team use cases. Led and developed the horse racing statistical analysis project using Knowledge Graph-based RAG with Neo4j and OpenAI. Implemented and improved two Turkish text classification models that serve four million active users with the BERT model. One is a binary classifier for customer comments, and the other is an issue classifier for customer complaints. Initiated, led, and delivered a customer service chatbot application that combines RASA for operational scenarios, OpenAI for unstructured scenarios, and an internal Qwen3 model for intent estimation. Developed an AI-based competition analysis for football matches using OpenAI and document-based RAG. Initiated, led, and delivered a personalization model based on a 2-tower algorithm. The project aims to recommend customer league and market-based matches and bets. Developed text classification models for the end user to classify their messages on the message board to eliminate malignant content. This BERT-based model achieved an F1 score of 0.96 for classifying such content. Extended the BERT model for intent classification of the customer messages that arrive at customer service. The 0.87 F-1 score helped the customer service agents reduce their workload. Improved and refactored the legacy models that forecast financial KPIs and anomalies. Introduced a money transfer optimization model across payment accounts. Also developed time-series projects to detect network and login anomalies. Initiated, led, and managed CRM projects for cost minimization and conversion maximization, including LTV, point distribution optimization, churn detection, and financial and behavioral segmentation models.

Developed data science and advanced analytics solutions to explore the use of patient data in the insurance domain. Introduced multiple models and their mixture to build explainable ML solutions for better insurance coverage. Implemented a claim lineage to build a knowledge graph using Neo4j to achieve that. Implemented an AI-based solution in the project's last phase, utilizing the previously explained knowledge graph. This RAG-based approach extracted necessary information from the KG and consulted the LLM for extra reasoning.

Created time series analytical modules to enable customers to monitor the performance of their PV-energy facilities. Designed and developed time series forecasting models for both power generation and power loss based on seasonal and irradiation data. Developed a comprehensive loss forecast model, coupled with a shading detection algorithm, to detect and classify loss amounts effectively. Established a streamlined data pipeline, optimizing data collection from MongoDB and Victoria Metrics into Redis for efficient utilization in analytical tasks.

Developed a B2B machine learning project, complete with an automation pipeline, to cater to diverse internal teams within the organization. Delivered a versatile API serving multiple stakeholders. Created automated Jupyter notebooks tailored to business stakeholders' needs. Introduced coverage metrics to assess the data pipeline's effectiveness and successfully unified disparate data sources. Orchestrated the automation of training, deployment, and model scoring processes in a containerized cloud environment, streamlining operations for increased efficiency.

Developed an analysis project powered by machine learning to estimate urban development and its effect on real estate at macro and micro levels. Augmented the dataset by collecting data from online resources, seamlessly integrating them into our ML model. Developed web scraping bots to collect data from online resources and websites efficiently and in an automated way. Developed user-friendly Jupyter notebooks and PowerBI dashboards to convey model outcomes and their business implications.

Implemented a customer lifetime value (CLV) prediction model based on loyalty points, contributing to the improvement in customer segmentation products. Worked on extracting loyalty KPIs and creating visualizations using Qlik Sense, collaborating closely with the business intelligence team. We built a data pipeline from the ground up using PySpark tailored specifically for analytical use cases. Collaborated with a consultancy team to deliver simulation projects and conduct in-depth analyses focused on exploring alternative strategies for increasing engagement. Developed, during the COVID-19 crisis, a recommendation tool geared toward increasing passenger engagement. This tool was centered around non-air loyalty items and utilized the ALS Library. Built a profile update-based fraud prediction and monitoring tool for loyalty

Developed predictive maintenance capabilities for thermal plants through a time series forecasting project focusing on FID fans within combustion engines. Collaborated closely with the engineering team at a thermic plant and successfully enhanced coal calorie prediction within the designated mining zone, leveraging the SGeMS tool and ordinary kriging techniques. Developed, in collaboration with the trading team, a model for forecasting the electricity market-clearing price. The outputs were instrumental in guiding the trading team's decisions regarding surplus and shortage pricing strategies. Created near real-time dashboards at both plant and portfolio levels within Grafana.

Took part in developing and deploying an online store on AWS by integrating the ERP services with the eCommerce platform. Undertook the implementation of a recommendation engine utilizing an ALS model. This innovative solution found its home on an AWS EMR instance. Collaborated with the eCommerce team to develop ad-hoc propensity models geared toward bolstering targeted marketing capabilities. Ventured into the development of back-end services for an in-house CRM tool. This endeavor allowed me to work extensively with NoSQL, particularly with Couchbase.