Yosuke Kuroki

- Quantitative Evaluation & LLM Research * Designed a novel GPT-4o-based evaluation framework for long-text generation and summarization, reducing feedback cycles from days to minutes and enabling rapid iteration. * Spearheaded R&D for "Client Intelligence," leveraging LLMs to extract structured insights from historical broker chatrooms and developing a text-to-SQL chatbot for natural language retrieval. - AI-Powered Lease Abstraction at Scale * Scaled an AI document processing platform by 100× (17k+ tenants/year), optimizing UI/UX, refining SME-driven prompt engineering, and doubling analyst efficiency—saving 8.5k+ hours annually. * Led a major system refactor (10k+ LOC), slashing per-tenant costs by 80% (from $25+ to $5) and reducing latency from 45 minutes to 5. - Technical Skills: * AI/ML: OpenAI, LLMs, LLMOps, Evaluations, Chatbots, OCR * Engineering: Python, Pydantic, Async Programming * Cloud/Infra: Azure, AWS

• Full-stack web development using .NET, JavaScript, and modern frameworks to build scalable, high-performance applications for enterprise clients. • Collaborate with cross-functional teams (product, UX, QA) to design, develop, and deploy end-to-end solutions that enhance user experience and business efficiency. • Optimize application performance through code refactoring, database tuning, and cloud integration (Azure/AWS). • Ensure maintainability and scalability by implementing clean architecture, RESTful APIs, and CI/CD pipelines.

• Spearheaded the creation of a hyperspectral tree species classifier, achieving a 30% increase in accuracy over traditional RGB methods, enhancing the company’s technological edge in environmental analytics. • Implemented a LiDAR-RGB fusion pipeline that significantly reduced false negatives in object detection by 30%, improving the reliability of data-driven decisions. • Redesigned the training infrastructure, reducing model development cycles from 7 days to 20 hours while cutting AWS costs by 80%, thereby maximizing resource efficiency and operational budget. • Co-developed a graph neural network architecture that halved the processing time for 3D point clouds, streamlining workflows and enhancing project turnaround times. • Streamlined annotation workflows, resulting in a 50-70% increase in labeling team efficiency, fostering a productive working environment and accelerating project delivery, while mentoring mid-level developers to enhance their skills and productivity.

• Developed a medicine ranking system leveraging Python, Databricks, and SparkXGBRanker, implementing EDA, feature engineering, and hyperparameter tuning to enhance model accuracy. • Managed cloud infrastructure on AWS EKS and Azure Kubernetes Service, provisioning resources via Terraform and Helm. • Designed and deployed scalable data pipelines on AWS S3/EC2 & Google Cloud Storage, optimizing preprocessing and model training for medical imaging datasets. • Enhanced model performance using data augmentation, transfer learning, and cross-validation techniques, improving robustness for real-world clinical applications. • Containerized ML applications using Docker, orchestrating secure, HIPAAcompliant deployments with Kubernetes to ensure real-time inference in clinical settings. • Deployed ranking models as UISE JVM Chassis applications, implementing realtime tracking, alerting, and periodic retraining, strengthening risk and customer engagement strategies. • Refactored and migrated legacy ML pipelines to Java & gRPC/OIPx protocol, modernizing a low-latency orchestrator for high-frequency trade execution and data consistency. • Established CI/CD pipelines on AWS & GCP cloud infrastructure, enabling seamless model updates and real-time risk reporting for financial analytics. • Designed & executed ETL pipelines using Pandas, PySpark, Jenkins, Airflow, Databricks, and Querybook, optimizing data ingestion & transformation workflows across cloud platforms. • Developed automated ETL workflows on AWS Lambda, EC2, and SageMaker, efficiently handling large-scale financial datasets. • Tuned Hadoop & Spark configurations, reducing processing time for critical big data operations, improving scalability & cost efficiency.

• Developed a CNN-based OCR Text Detection Model, reducing inference time for scanned documents and card images, optimizing processing speed and accuracy. • Designed and fine-tuned an OCR pipeline using Tesseract, digitizing and extracting textual metadata from archival labels, handwritten notes, and printed documentation. • Customized OCR models to recognize specialized fonts and handwriting styles in historical records, enhancing digital metadata enrichment. • Integrated denoising, text detection, and Tesseract OCR, leading to reduction in inference time for scanned certificate documents and PDFs. • Developed computer vision-based techniques for character & table detection in document scene images, leveraging OpenCV to enhance document processing efficiency. • Deployed distributed document processing pipelines on AWS (EC2, S3), ensuring scalability and secure high-volume storage of legal documents • Developed cloud-based RESTful APIs that seamlessly integrated ML models with legacy e-discovery systems, improving review turnaround time and operational efficiency. • Developed synthetic training data for logo detection, applying data augmentation techniques using OpenCV & C++ to improve model robustness • Deployed ranking models with Docker & Kubernetes as UISE JVM Chassis applications, implementing an orchestrator for performance tracking, alerting & retraining. • Developed a machine learning system for automated classification of physical media (tape formats, film reels, optical discs) via image and audio feature extraction

• Integrated real-time performance monitoring dashboards with Datadog, enhancing model observability and troubleshooting capabilities. • Designed automated ETL & preprocessing pipelines for legal & archival datasets, ensuring efficient data ingestion & transformation on AWS Lambda & SageMaker. • Conducted load testing with K6 & Locust, documenting findings & optimizing document processing pipelines for scalability & speed. • Designed automated ETL & preprocessing pipelines for legal & archival datasets, ensuring efficient data ingestion & transformation on AWS Lambda & SageMaker. • Designed & implemented an end-to-end NLP pipeline to automate the classification, review, and summarization of legal documents, reducing manual e-discovery time. • Collaborated with legal teams & compliance specialists to validate model fairness & transparency, ensuring ethical AI deployment in litigation support systems.