M.S. Artificial Intelligence candidate at Oregon State University, expected June 2027, with a B.S. in Applied Computer Science, AI focus. I build machine learning systems and research software, mostly on language-model agents and reinforcement learning, and I am applying to PhD programs in multi-agent language models.

• Portfolio: https://ancorro.github.io
• LinkedIn: https://linkedin.com/in/stevencm
• Personal email: steven.cleasby.mayeda@gmail.com
• University email: cleasbys@oregonstate.edu
• Location: Corvallis, OR

Seven of fourteen. The rest, with figures and write-ups, are at ancorro.github.io.

Isaac Lab, MuJoCo, PPO. Trained a Unitree G1 humanoid to walk to a table, grasp an object, carry it and set it down, and an SO-101 arm to pick and place from camera input alone. The learned grasp lifted and held 345 of 360 held-out objects, against 284 for the previous policy; camera-only placement reached 341 of 400 household objects. Negative results are kept beside the positive ones, and everything runs in simulation only.

Code: https://github.com/Ancorro/rlRobots

PyTorch, vLLM, Qwen3. Asked whether language-model agents can exchange internal state instead of text. A reproduction of LatentMAS on Qwen3-4B is bit-identical to the official code, and a split probe shows the content does cross: 91% with the right KV cache against 92% for text, and 0% with a wrong or absent cache. The latent team only ties a single agent, +0.1 and -0.2 points, so passing state does not beat one model doing the work. The latent sender is 19 times faster and uses 2.4 times fewer FLOPs.

Code: https://github.com/Ancorro/latent-comm-public · Paper draft in progress.

Python, tmux, Rust. Built a web app that runs on each of my computers and gives my phone a tab per agent session, with one-tap keys for Claude Code, Codex, OpenCode and Pi. Each agent runs in its own tmux session, so closing the phone or losing signal never stops it. Linked computers form one Deck with no main host, the phone app fails over after four seconds, and chats move between hosts mid-conversation. The server listens only on localhost and new devices pair with a one-time five-minute code.

Code: https://github.com/Ancorro/deck-public

WEC-Sim, PPO, Gymnasium. Wrote a Gymnasium interface that lets reinforcement learning agents step a full WEC-Sim wave simulation, matching a Simulink-only reference to about 1e-10 in mean power at 19.9 ms per control step. A PPO agent trained on a cheap linear model transferred with no loss: 148.8 kW on WEC-Sim against 147.4 kW for a tuned damper, trained in 2.4 minutes instead of 71.

OSWorld, Claude, Gemma 4. Built an agent that operates a Linux desktop by reading the screen into numbered items and choosing the next action. On 60 OSWorld multi-app tasks a plain single-loop design scored 27, ahead of the two-stage planner stacks at 16 and 20; the difference between those two is within noise at p = 0.42 by an exact McNemar test. The fast decision model answers in about 0.5 s per step against 5 s for the vision model.

Code: https://github.com/Ancorro/computer-use-stack-public

ESM3, LoRA, XGBoost. Predicted broadly neutralizing antibody breadth from heavy-chain sequence alone on 300 CATNAP antibodies. Frozen ESM3 embeddings raised AUROC from 0.651 to 0.773 over a handcrafted baseline, though the group-bootstrap interval includes zero. Validation held out all 109 metadata-derived lineages, and a categorical VAE on whole sequences placed 4 of 4 known VRC34 intermediates correctly.

Project page: https://ancorro.github.io/esm3-banb/

Gemini, Python, React. Won first of 20 teams in the beginner track at BeaverHacks 2026, building a multimodal agent in 24 hours that answers a technician's photo-and-voice question and draws the next step on the image, with QR-scoped retrieval over equipment manuals.

Code: https://github.com/ekanshArora123/beaverhacks2026 · Demo · Submission

Leading Software Engineer, Waste Treatment Completion Company (senior capstone, remote, Sep 2025 to Jun 2026). Led the student team building a Python desktop application for safety-critical Hanford radioactive-waste vitrification calculations. Architected validated calculation pipelines for auditability and numerical consistency, ran code reviews and task delegation, and wrote the developer handover package. Project page

Research Intern, Forever Sabah (Sabah, Malaysia, Jun to Sep 2023). Analyzed climate change in Borneo precipitation outliers from weather-model raster data in QGIS, and designed database schemas for wildlife image datasets supporting individual elephant re-identification.

Undergraduate Research Intern, Oregon State University (Jan to Jun 2023). Wrote data analysis scripts for research datasets, documented the methodology and presented the findings at the OSU undergraduate research conference.

• Languages: Python, C, JavaScript, SQL
• Machine learning: PyTorch, Hugging Face Transformers, transformer design, training and evaluation, ablation studies, LoRA and QLoRA fine-tuning
• Agents: agentic systems, multi-agent orchestration, multimodal pipelines, RAG
• Data: NumPy, pandas, statistics, feature engineering, data cleaning, Weights & Biases
• Systems: Linux, CUDA, Slurm and HPC, Docker, Git
• Web and backend: React, REST APIs, authentication, relational database design
• Practices: testing, CI/CD, code review, technical documentation

• M.S. Artificial Intelligence, Oregon State University. GPA 3.80, expected June 2027. 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 for 7 terms. June 2026.