Esma Nafiye Polat (PhD)

Multi-Agent CRM Meeting Assistant (LLM Orchestration, Enrichment, Routing)

• Designed an end-to-end agentic workflow for CRM activity logging (appointment/meeting/opportunity), including LLM orchestrator + deterministic graph routing and tool chaining.
• Implemented state model + missing-field control + follow-up logic, separating UI inputs vs system context (AccountNumber/PersonId etc.).
• Built enrichment pipeline (product/emotion/competitor extraction + summary payload build) with structured outputs and UI confirmation step.

Credit Card Statement Explanation Agent (Banking Decision Support)

• Developed an LLM-assisted agent to verify and explain credit card statement calculations (refunds, installments, overpayment, negative balance) for call-center decision support.
• Translated complex transaction tables into structured reasoning steps and customer-ready explanations while preventing hallucinations via controlled pipeline logic.

Web Content Ingestion PoC (Tavily Search API Evaluation)

• Investigated Tavily Search API for external web content ingestion to support knowledge base population and normalization workflows.
• Evaluated Tavily outputs across multiple websites for title/content quality, summarizable text extraction, and structured fields, documenting findings and limitations.

Agent Persona Detection as a Component of Virtual Bot Development

• Proposing and integrating additional persona attributes such as "Problem Solver" and "Empathizer" to improve customer satisfaction, operational efficiency, and revenue in telecommunications customer service environments.

• Generating and labeled over 6,480 synthetic conversations using GPT-4 with few-shot prompting to create balanced datasets, representing diverse agent personas and realistic conversational contexts. Synthetic data generation, model training, and evaluation were streamlined using the MLflow platform, enabling efficient tracking of experiments, data, and performance metrics.

• Developed LSTM and Transformer-based architectures for agent persona detection, leveraging embeddings like MPNet to classify personas based on attributes such as friendliness, verbosity, interruptiveness, and tone of voice. The MLflow platform was utilized to upload data, manage models, execute pipelines, and monitor key performance metrics, ensuring scalability and reproducibility throughout the process.

Mapping Free Text to Action List for Virtual Bot

• Developing virtual bot components for agent-caller interactions by spearheading the design and implementation of a system leveraging Large Language Models (LLMs) like LLaMA 3.1 and GPT-3.5 Turbo to automatically generate structured action lists from free-text instructions, enabling efficient configuration of Intents and Skills for admin users.

• Enhancing the adaptability and accuracy of LLaMA 3.1 and GPT-3.5 Turbo via few-shot fine-tuning, improving the model's ability to generalize from small, diverse datasets and handle complex instructions across multiple languages.

• Employing tools like Pydantic to validate model outputs and develop evaluation metrics, including cosine similarity and action prediction count matching, to ensure alignment between generated and reference action lists, facilitating accurate tool mapping.

Agent Profiling

• To increase Afiniti's revenue and recommendation engine effectiveness by by 0.8% generating impactful NLP features and developing advanced reranking systems.

• Conducted research on agent profiling to identify impactful features for detecting general agent behaviors using agent-caller transcripts and agent summary text data.

• Developed agent profiling models by extracting features from agent summary data utilizing machine learning, transformer-based, and LLM-based models, assessing their impact on the recommendation engine and driving a 0.8% revenue improvement.

• Implemented detection of offer acceptance using MPNet and Llama 3.1 embeddings, and client intent prediction using the Mistral model derived from agent-caller interactions.

• Integrated agent-based features from both agent-caller interactions and summary data into comprehensive agent profiling models using advanced machine learning and LLM approaches, including Zero-Shot and Few-Shot Learning techniques.

• Preparing to apply for a patent on these innovative agent profiling techniques and their integration with NLP-driven recommendation systems.

Hamilton Project

• To explore and integrate advanced LLM models into Afiniti’s system for enhanced data processing and actionable insights.

• Utilized transformer and LLM-based algorithms, including LexRank, TextRank, BART, T5, and Llama 2, for summarization applications on call transcript data, enabling efficient distillation of long conversations.

• Employed topic detection methods such as LDA, BERTopic, and advanced LLM-based approaches to extract actionable insights from call transcripts.

• Focused on emotion detection in call transcript utterances using Hugging Face-based models, EmoRoBERTa, and FastBERT, enhancing the system’s emotional intelligence and understanding of customer-agent interactions.

Chatbot Development Optimization

• To develop a marketable virtual bot for agent-caller interactions, improving customer satisfaction and reducing operational costs.

• Identified and analyzed agent and customer behaviors to enhance the functionality of a newly developed chatbot, focusing on improving customer satisfaction and efficiency in agent-caller interactions.

• Investigated customer and agent behaviors related to churn, leveraging insights to optimize the chatbot’s effectiveness.

• Applied advanced LLM techniques, including Few-Shot and Zero-Shot Learning, to improve the chatbot’s adaptability to diverse conversational scenarios and its ability to handle new tasks with minimal retraining.

• Pioneered the creation of the "Couple Dialogue" dataset, the first dedicated to conversational sentiment analysis in Turkish, with comprehensive statistical validation.

• Conducted the inaugural study applying contextual and non-contextual modeling techniques to Turkish conversational sentiment analysis, demonstrating significant enhancements.

• Utilized advanced Large Language Models (LLMs) including GPT-4, Llama 2, and GPT-3.5 Turbo, marking a novel application for Turkish sentiment analysis.

• Achieved a breakthrough 31.12% increase in Weighted F1 scores using fine-tuned models over prompt-based approaches, validated with paired t-tests.Identified and integrated specific Turkish morphemes into sentiment analysis, achieving up to 9.89% improvement in classification accuracy across diverse settings.

• Demonstrated that BERT-based utterance features outperform traditional vector embeddings, with substantial F1 score improvements in both model frameworks.

• Developed a robust baseline with non-contextual models employing a variety of embeddings such as BoW, TF-IDF, and Word2Vec, providing a comprehensive comparison against contextual models.

• Led a study that established and rigorously tested the superior performance of contextual models over non-contextual models in Turkish, setting new standards in the field.

• Conducting research on leveraging GPU acceleration and efficient parallelization techniques to enhance computational efficiency in a PhD thesis focused on conversational sentiment analysis, while also exploring the application of advanced transformer models and LLMs to improve contextual frameworks and achieve superior performance.

• Estimating POS Terminal revenue and POS Terminal potential revenue for bank’s POS customers.

• Estimating the top credit card limit for private bank customers to increase credit card usage and maximize the bank profit. Increased company profit by 5% in less than 6 months.

• Forecasting customer bank cash flow to project balance sheet items accurately resulting in savings of 100M TRY.

• Predicting the attachment styles of recently married couples using acoustic features such as acoustic features and I-Vectors on couple interaction data.

• Reducing uncollectible accounts by 20% by predicting top credit limit, top credit cash and non-cash limit, top product limit, and the top collateral limit for small and medium-sized enterprises.

• Doing branch optimization to provide efficiency and customer satisfaction by 30%.

• Determination of fraud customer checks at an early stage.

• Generating reports which show daily price in two distinct ways, namely minimum price and minimum connection price

• Generating automatic alert system based on user criteria.

• Combining different data sources like airline's market share, reservation information, load factor with airline's price information

• Generating making inference system.

• Generating user action system.

• Estimating Turkish Airline’s price values.

• Estimating Turkish Airline’s load factor values.

• Estimating Turkish Airline’s price elasticity values.

• Guiding students in the lab, grading lab reports, homework, and exams in Java Programming Language course.

• Doing weekly recitation for student groups, preparing quiz questions and their solutions, and organizing office hours in courses of Python Programming I, Python Programming II, Systems Fundamentals, and Unix Operating Systems.