Kevin Takano

Built a production LLM evaluation workflow for rent inspections and an LLM-as-judge feedback loop that improved classification quality by 10%. Analyzed internal AI-agent architecture and built multi-step conversation scoring pipelines, improving quality score from 0.48 to 0.90.

Built an FAQ agent for medical appointment scheduling, improving response quality by 10–15% and cutting latency in half. Developed an end-to-end CatBoost conversion model reducing costs by 50% and built Twitter and Telegram LLM agents with ElizaOS plugins.

Built a Scala and Spark calibration model for credit-card interest-rate prediction, reducing incorrect predictions by 80%. Improved engineering velocity with reusable libraries, unit tests, and faster CI/CD, and connected model behavior to business outcomes through automated reporting pipelines.

Led an end-to-end embedding recommender system on SageMaker and Kubernetes, increasing click conversion to 8.6% versus a 2.07% baseline and driving 3.8x more purchases. Built a treasury forecasting system for PIX and billet payments that reduced error by 5% and re-architected chatbot context and document-processing flows.

Built dashboards, supervised and unsupervised models, and BigQuery pipelines for digital-bank acquisition; clustering reached 5% conversion versus 1% from random targeting.

Built Scala and Spark pipelines for segmentation, identity resolution, and recommendations; reduced AWS ingestion jobs from hours to minutes with incremental processing.

Developed Java, C++, and Node.js components for Android and Ginga middleware, plus CI and test automation that cut test runtime from hours to seconds and reduced library size by 50%.

Master's in Informatics with a focus on Machine Learning, completed in 2022.

Bachelor of Computer Science, completed in 2016.