Dogugun Ozkaya

Initiated, led, and delivered a customer service chatbot application that can handle structured scenarios, free text conversations, and intent classification.

In this complex structure, an intent classification model utilizing a few-shot approach with a local Qwen3-235B model was placed.

Free text questions for unstructured scenarios were handled with Agentic RAG using Elasticsearch with RRF, where retrieved documents were rephrased with OpenAI. Another tool was designed to fetch user history from Redis to ensure the relevancy of the user messages. Also in the final phase, I introduced Langfuse to monitor and track user questions and the relevancy of responses.

For the structured scenario use cases, especially in financial transactions, I used the RASA framework. In this phase, I managed the user story with predefined buttons and fixed user input, e.g., amounts and IBANs.

This multi-layer solution was designed to improve the medical insurance claims after they are rejected, so that they will be paid back.

The initial step was to determine if the claim would be rejected with C4.5 decision trees and, if so, what factors are affecting it using the pathways of the DT.

Then we focused on building the claim lineage to determine the story of the patients and their claims.

Finally, we utilized Claude's API to determine the root cause of the claim's response. To better resolve the underlying ICD 10 terms, we built a RAG system from the claims textbooks and and documentation about ICD 10. We processed the books and documents with Docling and used them in an RAG system utilizing FAISS.

As the lead AI engineer, I managed and implemented the project of generating horse race-specific statistical analyses.

The knowledge base was built on Neo4j, and the Cypher query generation was provided by the tool's use of the OpenAI API. The generated statistical information in JSON format was rephrased using the OpenAI chat completion API.

The knowledge graph contains global horse racing information, including all the details: races, horses, jockeys, trainers, training, and so on. The generated statistical analysis content provides helpful facts for betters and domain experts.

Designed and developed a text classification system using Google BERT's specific Turkish model to classify customer texts in two different tasks.

One is the classification of user messages to a messaging board for a betting/sports company. The model is trained monthly with a local GPU cluster. The deployed model is running on a 4-node cluster and serving up to three million active customers daily. The current F1 score of the model is 0.96.

Another version of the same model is trained to specifically detect the customer intent with the message he sent to customer services. This successful model, with an F1 score of 0.87, significantly reduced the effort required for customer service agents to manually select the topic.

The models are managed by MLFlow and served with FastAPI.

Developed and managed a customer analytics product for the customer service team, where the project included: GDPR data masking, sentiment analysis, intent classification, and speech-to-text transcription components.

To complete GDPR masking, we utilized regex patterns and a Turkish NER model from Huggingface.

For the sentiment analysis part, for both intensity (negative-neutral-positive) and category (happy, angry, sad), we utilized a few-shot methodology with the local Qwen3 model.

The intent classification was implemented using the legacy BERT model, which detects the topic of the customer complaint.

To include call center data in the solution, we used the whisper-large-v3 model from Huggingface, then summarized it with the local Qwen3 model before sending it to the pipeline of other components.

An industrialized B2B machine learning solution. In this engagement with Toptal, I was part of a team focused on developing the solution.

My role encompassed the creation of the ML model and an API to serve real-time needs within the organization. Additionally, I visualized coverage metrics within the data pipeline, optimizing the process. In the final stages of the project, I worked on visualizing the data model outputs and SHAP values. I contributed by building an MLOps pipeline and automating the training, deployment, and scoring processes of the ML application within a Dockerized environment on AWS.

An analysis project powered by machine learning solutions to estimate urban development. The initial step was to determine the KPIs to use as the target variable. I scraped websites like The Good Schools and Walk Score with a user-like and efficient approach. I also developed Jupyter notebooks and PowerBI dashboards to deliver model outputs.

In this chatbot project, we aimed to develop a chatbot that utilizes state-of-the-art LLM models to deliver concise information from drug package inserts.

For this purpose, we captured drug labels from the FDA's web resources in XML format and converted them to PDF. Then, we utilized LangChain and HF embeddings to convert them to vectors and upload them to the designated Pinecone instance.

In the final phase, we employed OpenOrca's Mistral model to generate answers from the list of documents we captured from our vector database.

The end solution is deployed as an API with Flask to be served on an AWS EC2 instance.

A demo showcase project that imitates my work in the power generation industry to estimate the value of the sensor in a 12-hour prediction horizon from its and other related sensor values in a 6-hour time window.

In the original project, the goal was to estimate the vibration of the combustion fan. The input features were temperature, dust collection, and airstream sensors. The target value is the vibration frequency in 6-9-12 hours of prediction horizons.

The input data is collected in SQL Server from Osisoft PI's IoT data. The model is deployed to the on-premise server, and the outputs are visualized in Grafana, along with other input feature values.

In the original project, a GBM was used for retraining and retuning easiness. This demo version uses LSTM as an alternative approach.

Finally, the deployed model and predictive monitoring dashboard are used by supervisor shifts to plan their maintenance program proactively.

In this challenging task, I aimed to predict the market-clearing price for the energy market. The output of this model is used to plan the week ahead surplus and shortage energy price possibilities.

The model data consisted of renewable forecasts, demand forecasts, plant availability declarations, and natural gas prices. The data is collected in a DWH model in SQL Server from APIs, web scraping, and Excel file reports.

A model is developed with gradient gradient-boosting regressor from the scikit-learn library.

I developed a recommendation engine for mavi.com, an online store, using PySpark's ALS model. We used purchase data and access logs to create personalized product suggestions, aiming to boost the conversion rate and introduce cross-category recommendations. The project successfully improved user engagement and conversion rates, highlighting the power of data-driven strategies in eCommerce.

In my master's thesis, supervised by Prof. Albert Guvenis, I conducted a comparative analysis of blood glucose prediction models. We focused on replicating existing models using AIDA2 simulator data, categorizing them into two groups: non-parametric and neural networks.

Within the non-parametric category, we compared Random Forest Regression (RFR), Gradient Boosting Machines (GBM), and Support Vector Machines (SVM). We pitted Long Short-Term Memory Networks (LSTMs) against Adaptive Neuro-Fuzzy Inference Systems (ANFIS) in the neural network category.

The study revealed that SVM outperformed in terms of Root Mean Square Error (RMSE) among the non-parametric models, while ANFIS demonstrated superior performance in neural networks, surpassing SVM.

This project uses a web scraper to capture profile and activity information from top-level social media accounts for compliance reporting.

The automatized scraper is triggered by corn jobs daily, and the collected data is saved on PostgreSQL DB in AWS.

In a personal finance project, I created a Python script to collect data from finance.yahoo.com for a selection of Turkish stock market companies. The script retrieves data on EPS, 52-week range, and quarterly earnings growth for each company, then ranks the data based on these criteria. The results are saved and emailed.

Technically, the script runs on my personal AWS EC2 cluster. A Lambda service initiates the EC2 instance, and once the script completes its run, it shuts down the instance. The resulting files are archived in an S3 bucket.