Steven Cleasby-Mayeda

AI / ML Engineer

Oregon State University Corvallis, Oregon

M.S. in Artificial Intelligence candidate (expected June 2027), with a B.S. in Applied Computer Science (AI focus). I build machine learning systems and research software for ML and biological research, with an emphasis on reliable implementation, evaluation, testing and documentation.

Research interest: LLM research. Applying to PhD programs in this area.

StackPyTorch · LLMs · Agents ResearchLLM research StatusApplying to PhD programs

Projects

Research, systems and games · newest first · click any card for details
Simulated Unitree G1 humanoid picks a small yellow box off one table, walks to a second table and sets it down
Personal Research · Robotics · RL · ML

Humanoid and Robot-Arm Manipulation in Simulation

Isaac Lab · MuJoCo · PPO
Sep 2026 – present · 4 videos · 1 imageIn progress

A simulated Unitree G1 humanoid walks, grasps, carries and places objects; its learned grasp picks up 345 of 360 unseen objects, up from 284.

Details Code
Diagram: (a) four workers send long text reports to a manager; (b) workers send internal-state summaries and a manager routes them only to the workers that need them, with some not sent under a budget.
Doctoral Research Direction · Multi-Agent Language Models

Learned Routing of Latent State Between LLM Agents

PyTorch · vLLM · Qwen3
Oct 2026 – present · 5 imagesIn progress

Can LLM agents share internal state instead of text? Passing KV caches does carry information (91% vs 92% for text), but does not beat a single agent.

Details CodePaper draft in progress
Bar chart of mean absorbed power by seed for two fixed dampers and two PPO agents, evaluated on WEC-Sim and on the state-space model
Graduate Project · Oregon State University · Energy Systems

Reinforcement Learning Control for a Wave Energy Converter

WEC-Sim · PPO · Gymnasium
Oct 2026 – present · 3 imagesIn progress

A Python interface that lets RL (reinforcement learning) agents step a full WEC-Sim wave simulation; a PPO agent trained on a cheap linear model transferred with no loss.

Details
Diagram of the agent loop: screen read into numbered items, a fast decision model acts when confident, otherwise a vision model picks the action; kept notes feed back each step.
Personal Research · Language-Model Agents

Computer-Use Agent Evaluated on OSWorld

OSWorld · Claude · Gemma 4
Oct 2026 – present · 2 imagesIn progress

An agent that operates a Linux desktop; on 60 OSWorld (desktop task benchmark) tasks, a simple single-loop design scored 27, ahead of the planner stack's 16 and 20.

Details Code
Paired bar chart of the share of held-out lies removed by each unit set and by its best random control, for Qwen3.5 0.8B, 2B and 4B.
Personal Research · Interpretability

Do Small Models Have Honesty Heads?

PyTorch · Qwen3.5 · Steering
Oct 2026 · 2 images

A replication of honesty-head steering on three small Qwen3.5 models; a null result, since the one large effect just pushed the model to say Yes.

Details
Diagram: my finished writing, split into passages, rewritten generically by a base model, then QLoRA learns to map the generic text back to my original, scored by chrF, held-out loss and an AI detector.
Personal Project · Fine-Tuning

Rewriting Text in My Own Writing Style

QLoRA · DPO · vLLM
Oct 2026 – present · 2 imagesIn progress

A QLoRA fine-tune that rewrites text in my style from ~20,000 words of my writing; it nudged chrF from 65.0 to 67.5, while preference tuning only memorized.

Details Code
ROC and precision-recall curves for the antibody breadth classifiers
Graduate Course Project · Protein Language Models

AI Modeling for HIV Antibody Research

ESM3 · LoRA · XGBoost
Jun 2026 – Sep 2026 · 4 images

Predicting HIV antibody breadth from sequence: frozen ESM3 embeddings raised AUROC from 0.651 to 0.773 on 300 antibodies, though the interval includes zero.

Details Project page
Two ECG traces of record 222: the raw one sags with baseline drift; the filtered one sits level with every beat intact.
Personal Project · Biomedical Signal Processing

ECG Heartbeat Classifier on MIT-BIH

Python · SciPy · Signal processing
Oct 2026 · 3 imagesIn progress

An ECG (electrocardiogram) heartbeat pipeline on MIT-BIH; so far it loads all 48 records (109,494 beats) and filters out 56.4 dB of powerline hum.

Details

Experience & Research

Leading Software Engineer (Senior Capstone Project)
Waste Treatment Completion Company, Richland, WA. (Remote)
Sep 2025 – June 2026

Lead technical developer for a Python application processing safety-critical calculations for Hanford radioactive waste vitrification. Developed data-analysis pipelines focused on auditability, consistency, and maintainability; documented the implementation and created the developer handover document; managed the student team through task delegation and conflict resolution.

Research Intern
Forever Sabah, Sabah, Malaysia
Jun 2023 – Sep 2023

Analyzed climate change impact on Borneo precipitation outliers using weather model raster data and QGIS. Curated wildlife image datasets and established database schemas for individual elephant re-identification.

Undergraduate Research Intern
Oregon State University, Corvallis, OR
Jan 2023 – Jun 2023

Developed data analysis scripts to process research datasets and highlight trend patterns. Documented methodology and presented research findings at the OSU undergraduate conference.

Education

M.S. Artificial Intelligence
Oregon State University · GPA 3.8
Expected June 2027

Relevant Coursework: Foundation Models, Deep Learning, Optimization for AI, Data Cleaning for ML, Graph Theory, Operating Systems II.

B.S. Applied Computer Science (AI Focus)
Oregon State University · GPA 3.37 · Major GPA 3.52 · Honor Roll (7 terms)
June 2026

Relevant Coursework: Web Development, Databases, Software Engineering, Algorithms, Machine Learning & Data Mining, Mathematical Statistics, Parallel Programming, Reverse Engineering Malware.

Current Research

LLM Research

LLM Agents · Reinforcement Learning

Studying how teams of language models should communicate, including a proposal for learned peer selection over agents' latent states. Also starting reinforcement learning experiments for robotic grasping.

Skills

Languages
Python, C, JavaScript, SQL
Web & Backend
React/MERN, REST APIs, authentication, relational database design
ML
PyTorch, HuggingFace Transformers, transformer design, training/evaluation, ablation studies
Agents
Agentic systems, multimodal pipelines, RAG
Data
NumPy, optimization, statistics, feature engineering, data cleaning/visualization
Systems
Linux, CUDA, Docker, Git/GitHub
Practices
Testing, CI/CD, code review, technical documentation