Berkay Tor

Led design and implementation of scalable NLP data pipelines on Databricks for web page classification, sentiment analysis, and entity recognition using PostgreSQL for storage. Developed and fine-tuned transformer-based NLP models on AWS to improve content relevance prediction, increasing classification accuracy to >90% and driving a 15% uplift in ad click-through rate. Utilized MLflow and Databricks notebooks for experiment tracking and reproducible workflows. Deployed FastAPI microservices on AWS ECS/EKS with Redis caching supporting sub-100ms inference latency for real-time ad serving. Collaborated cross-functionally to integrate NLP-derived signals improving ad relevance and campaign ROI while monitoring performance using AWS CloudWatch.

Architected and implemented MappedGPT, an LLM-powered query interface using FastAPI and GraphQL integrated with OpenAI GPT-4 for translating natural language to GraphQL queries, reducing query development time by ~50%. Developed retrieval pipelines using PGVector and Haystack for retrieval-augmented generation with sub-second latency and better relevance. Built a conversational UI with LibreChat integrated into the GraphQL console, boosting usage by ~30%. Containerized microservices using gRPC deployed on Kubernetes with auto-scaling, sustaining thousands of concurrent sessions with 99.9% uptime. Designed and implemented Anaxi, a smart multi-tool AI chatbot powered by GPT-4 and the Model Context Protocol for dynamic invocation of domain-specific tools, improving answer accuracy by ~15%. Engineered full-stack infrastructure with FastAPI, gRPC, Docker, and Kubernetes delivering sub-300ms median response times.

Ensured HIPAA compliance by implementing encrypted data pipelines and secure storage for protected health information. Developed TensorFlow CNN computer vision models analyzing medical imaging with ~90% accuracy to support clinical decision-making. Built NLP workflows using BERT-based transformers for extracting insights from electronic health records improving patient risk stratification. Engineered full-stack applications with React.js front-end and Node.js/Flask back-end integrating AI predictions and real-time visualizations. Deployed AI services securely on AWS and Azure with containerization and CI pipelines for scalable production deployment. Collaborated with clinicians to validate models and ensure healthcare standards compliance.