Yavuz Sunor

• Led the architecture design and deployment of 2 main recommendation and ranking systems, including retrieval and prediction models, enabling relevant content discovery via natural language queries. • Designed, productionized, and maintained scalable APIs for search and discovery tools, collaborating with Product Managers across Content Planning, Marketing, and Ads teams to support key business objectives. • Designed and deployed complex multi-agent systems (agentic workflows) using Google ADK & A2A and the LangChain Agents library, leveraging multimodal content (scripts-subtitles + video assets) data. • Served as a technical lead, mentoring engineers on best practices for designing and scaling agentic systems and distilling complex system designs for executive review and roadmap alignment. • Fine-tuned foundational LLMs on movie screenplays to predict narrative dynamics and character-driven features, providing insights for content programming and acquisition decisions. • Built scalable batch and online data pipelines that integrate internal and 3rd-party APIs into downstream APIs and microservices supporting recommendation and search. • Improved system reliability and responsible deployment by designing and implementing robust monitoring, alerting, and A/B experimentation pipelines for high-traffic LLM and multi-agent services. • Drove end-to-end ML development: from ideation and proof-of-concepts to productionization with AWS (EC2, S3, Lambda), Airflow, SQL/noSQL DBs, and CI/CD pipelines using GitHub, Docker, and Kubernetes.

• Developed and deployed a semantic recommender system for merchant reviews, utilizing an open-source LLM (Llama) and embedding model (Sentence Transformers) to improve personalized merchant discovery and ranking. • Drove the end-to-end ML pipeline for a new credit card product, leading POC creation and deploying LLMs in production to enhance both customer experience and credit risk evaluation. • Modeled credit card transaction behavior with BiLSTM + Attention for a no-preset limit credit card, optimizing credit allocation. • Prototyped and deployed customer support chatbot using LangChain, GPT, and RAG with financial transaction data. • Built OOP-based data pipelines in Python with PySpark and Snowflake, processing real-time transaction streams at million-row scale and performing systematic feature engineering for high-quality, high-velocity inputs to downstream ML models. • Deployed ML pipelines and models in AWS cloud environments with Docker and optimized inference using ensemble methods (XGBoost, GBM) and neural networks (PyTorch/Keras).

• Built a series of data pipelines, designed a variety of solutions(from simple heuristics to ensemble trees and neural networks) for early fraud detection during sign-up processes of neobanks. • Built feature engineering classes and methods to encode unstructured texts and JSON payloads from real-time user sign-up forms and 3rd party vendor APIs • Implemented real-time and batch model components of whole data science pipeline in Python, building and deploying Lambda functions in AWS, creating ETL pipelines from RDS to Redshift using Glue • Applied Word2vec and TF-IDF vector transformations using NLTK and Gensim libraries on unstructured 3rd party vendors data for classifying documents from each source and calculating similarities among them • Trained, tested and deployed machine learning models including Gradient Boosting Trees, Random Forests and Neural Nets for fraud prediction and generated synthetic data with GANs for data augmentation • Models in production provided 75% drop in rejection rates in sign-up processes and reduced the fraud rates from 5% to 0.5%

• Designed and implemented feature engineering pipelines for default probability prediction using demographics, past credit data, occupation and social network features • Trained, tuned and tested several machine learning models including Logistic Regression, k-NN, SVM, GBMs for default probability prediction and unsupervised models for user clustering including k-Means and Gaussian Mixture model • Wrote advanced SQL queries on clients’ profiles and financials to create monthly and ad-hoc dashboards • Built autoregressive(ARIMA) time series models using macroeconomic data in R for banking stress testing • Developed Customer Life Cycle models in R incorporating customer's financials and risk profiles

Coursework highlights:​ Deep Learning(​Yann LeCun)​, Machine Learning, Computational Cognitive Modeling, Data Visualization Projects:​ • Combined unsupervised learning and LSTM time-series models to predict emerging crime hotspots in NYC • Gauged bus ridership at a chosen bus-stop with OpenCV image processing using camera footage • Implemented ResNet architecture for Image Classification applying data augmentation and self-supervised learning • Developed interactive visualizations and web apps using Javascript, D3 & Vega-Lite

Coursework highlights: Calculus, Linear Algebra, C Programming, Probability and Statistics, Numerical Analysis