Skip to content
HomeProjectsHighlightsBlogPlaygroundRésuméAbout

About Me

01 / Background

I am Michael Rusu, a Computer Science and Data Science undergraduate at the University of Central Florida’s Burnett Honors College, based in Orlando, Florida. I expect to graduate in June 2028.

I’m an undergraduate researcher at UCF Research Labs and a former Machine Learning Intern at Lockheed Martin. My work spans distributed model training, local language-model systems, and machine learning for aerospace inspection.

Michael leading a technical workshop for students
Teaching and building with the UCF developer community

02 / Research

At UCF Research Labs, I work on model distillation and distributed fine-tuning: helping smaller models learn from larger ones and scaling training across multiple GPUs. I’m also studying why large language models hallucinate.

I’m about to start research for the DoD, where I’ll be working on reinforcement learning and machine learning.

I also wrapped up my KRSP research on local GraphRAG systems, bringing together language models, embeddings, and knowledge graphs, and presented a poster on the work. You can check out my projects or read my technical notes for more.

03 / Hobbies

When I’m not training models, I’m usually playing guitar, hiking, rock climbing, listening to music, or experimenting with hardware.

Michael playing guitar on stage
Playing guitar
Michael snowboarding on a snowy mountain
Getting away from the screen

04 / Experience

Internships

Machine Learning Intern

Lockheed Martin

Summer 2024 · Tampa, FL

Built distributed data pipelines, trained geometry models, and deployed containerized inference services for CAD comparison and aerospace inspection.

  • Built a distributed ETL pipeline designed to ingest CAD models, revision metadata, and inspection outcomes. Applied schema validation, deduplication, and geometric normalization to produce training datasets.
  • Trained a Siamese geometry encoder across multiple GPUs using PyTorch DDP, mixed precision, distributed sampling, and contrastive loss. Fused learned embeddings with FPFH feature matching, RANSAC–ICP registration, and Chamfer/Hausdorff distance for CAD comparison.
  • Containerized inference services with Docker and Kubeflow, delivering an 86% improvement, 58% faster inspections, and $15,000 in weekly savings. Validated models to under 5% error across 100+ tests with 10+ engineers across QA and operations.

MIS Intern

SCC Soft Computer

Summer 2023 · Largo, FL

Supported network infrastructure and security for healthcare systems serving hospitals and laboratories.

  • Managed 200+ server-port configurations and optimized network pathways.
  • Implemented security measures for 100+ network ports.
  • Reimaged or repaired 50+ systems while maintaining hardware inventory.

Research

Undergraduate Researcher

UCF Research Labs

Aug 2025 – Present · Orlando, FL

I work on model distillation, distributed fine-tuning, and hallucinations in large language models. I’m also about to start research for the DoD, working on reinforcement learning and machine learning. I recently wrapped up my KRSP research on local GraphRAG systems and presented a poster on the work.

  • Distilled Qwen2.5-7B through teacher–student supervision and scaled distributed fine-tuning with MPT across four GPUs using DDP, BF16, gradient checkpointing, and sharded dataloading. Cut training from 41 to 19 hours while retaining 93% of teacher performance on 600 held-out prompts.
  • Built an instrumented 3D Unity environment for fine-tuning action-conditioned world models. Reduced held-out rollout error by 14%, measured using pose-transition RMSE and perceptual similarity.
  • Built a local GraphRAG stack with 4-bit MLC-LLM inference, MiniLM embeddings, LanceDB, and four query modes. Executed TVM/MLC source builds, quantization, and Metal/CUDA compilation, plus resumable 16-worker graph extraction with cloud-teacher targets for Ministral-3B supervised fine-tuning.

Ambassadors

Codex Ambassador

OpenAI

2026 – Present · Remote

Supporting the global developer community with practical Codex education, agentic coding workflows, and product feedback.

  • Demonstrate Codex workflows for building, debugging, and shipping software with coding agents.
  • Create developer education that makes agentic coding patterns practical and responsible.
  • Share community feedback that helps improve how developers learn and work with Codex.

Comet Ambassador

Perplexity AI

Jan 2026 – Jun 2026 · UCF

Introduced Perplexity Comet browsing and AI-assisted research workflows to the UCF community during a completed six-month ambassadorship.

  • Demonstrated Comet browsing, search, and research workflows to students.
  • Organized UCF demos and helped students and faculty evaluate AI-assisted web research.

Leadership

Director of Workshops & Hackathon Organizer

Knight Hacks

Aug 2025 – Present · Orlando, FL

Leading technical workshops and organizing hackathon programming for the UCF developer community.

  • Organized and led five technical workshops reaching 300+ participants.
  • Built cross-club partnerships and collaborated with external organizations including OpenAI and Hugging Face.
  • Coordinate workshop programming, event logistics, and participant engagement.

Project Lead

AI Club ELHS

2023 – Jun 2025 · East Lake High School

Led student machine-learning projects and taught AI fundamentals through June 2025.

  • Led a team of 15+ members building machine-learning applications and research projects.
  • Mentored students in deep learning, computer vision, NLP, reinforcement learning, and robotics.
  • Presented work at school and district events.

05 / Writing

271,104 Routes: What I Missed About Mixture-of-Experts

August 2026 · 17 min read

how can a model perform better if it only uses some of its experts? i wanted to understand how that worked, so i tried implementing the pieces in pytorch and looked at 271,104 routes from granite.

06 / Contact

I’m open to internships, research, and ambitious collaborations. I’d love to hear what you’re building and where I can help.

mickirusu@gmail.com

X / @reach_mrpieGitHubLinkedInRésumé

with by michael