Michael E. Marin

Mr. Cooper Group — Dallas, TX

• •Contributed to development of enterprise agentic AI products automating customer service workflows for one of the largest mortgage servicers in the U.S., supporting millions of customer interactions annually.
• •Authored user stories and feature specifications in Azure DevOps for ML-powered automation systems, translating business requirements into development-ready tasks aligned with sprint cycles.
• •Executed model evaluation cycles measuring accuracy, reliability, and production performance of AI outputs against defined benchmarks; flagged regressions and coordinated fixes with engineering.
• •Facilitated user testing sessions with internal stakeholders, synthesizing feedback into prioritized backlog items that improved feature adoption and reduced UX friction.
• •Coordinated cross-functional delivery across engineering, design, and operations teams using GCP, MySQL, Microsoft Copilot, and LucidChart to support digital transformation initiatives.

• •Improved Recall@10 from 58% to 81% by replacing dense-only retrieval with a hybrid pipeline combining pgvector ANN search and BM25 lexical matching, followed by cross-encoder reranking.
• •Reduced irrelevant context rate from 31% to 11% through semantic chunking and reranking, increasing answer precision by 35% on a 150-query human-labeled benchmark.
• •Designed asynchronous document ingestion pipeline (PDF, Markdown, HTML) using FastAPI, PostgreSQL with pgvector, Redis, and Celery; supports knowledge bases of 1,000+ documents.
• •Implemented per-query observability layer capturing retrieval latency, token usage, and cost per request; surfaced via admin dashboard, enabling targeted optimization without re-running experiments.
• •Built evaluation framework with 150+ benchmark queries measuring Recall@k, MRR, and answer correctness; used to quantify impact of every pipeline change before deployment.

In collaboration with NASA Goddard, University of Maryland, and UNT

• •Led team of 5 engineers to build a real-time interactive data visualization platform for NASA's PACE satellite mission, enabling researchers to explore live oceanographic and atmospheric data streams.
• •Architected full-stack cloud data pipeline in React and TypeScript, processing large-scale multi-channel satellite datasets with sub-second render latency for scientific analysis.
• •Designed Setup Wizard using ImGui and OpenCV, automating researcher environment configuration and eliminating a multi-step manual process prone to errors.
• •Built GraphQL API layer integrating multi-source NASA scientific datasets across platforms, enabling consistent data access for downstream analysis tools.
• •Selected for public exhibition at Kennedy Space Center Museum, Washington D.C.; project scheduled for national museum tour to bring PACE mission science to general audiences.