Projects

Production test automation at Real Time Automation and a full-stack deep RL training platform.

May 2025 – Mar 2026

MSOE senior design

TODO: short screen recording or GIF of training and live playback

RL Studio

An educational web app for configuring and launching Atari agent training with DQN, Rainbow, and PPO, playing the same games live in the browser, and viewing training analytics.

  • FastAPI backend with REST endpoints to create, reset, step, and render Gymnasium environments, plus a WebSocket stream for live frames and rewards during training.
  • Config-driven training service wrapping Stable-Baselines3, with user-configurable hyperparameters, eval/checkpoint callbacks, and JSON progress logging.
  • Next.js + React + TypeScript + Tailwind frontend with guided hyperparameter tuning, real-time training charts, and live game playback; GitHub Actions CI for ESLint, Prettier, and TypeScript checks.

I chose RL Studio for my team because reinforcement learning is showing up in almost every field, yet its jargon scares a lot of people away before they ever try it. We built a platform where anyone, technical or not, can learn by doing: configure hyperparameters, train real agents, and get help from a built-in LLM assistant and plain-language pop-ups for every term.

TODO: link to repo or demo


May 2025 – May 2026

Software Engineer Intern, Real Time Automation

TODO: a screenshot of an Allure trend report with no company data, or a diagram of the nightly flow

Automated regression pipeline

Designed an unattended nightly pipeline (crontab + Allure) running 300+ regression tests across 7 industrial protocols and publishing trend reports to engineers' network drives, viewable in Azure DevOps, replacing manual test runs for 30+ engineers and enabling failure analysis across releases.

  • Wrote the majority of the regression suite (MQTT, HTTP/S, SQL, FTP, WebSockets, email, Modbus, PLC read/write, Selenium front-end checks) for the company's IoT Historian product; took part in weekly code review on Azure DevOps.

The Historian had 300+ tests and a full run took over six hours, so every code change meant a long wait to find out whether it broke something. I built a pipeline that runs the full suite overnight (8 pm to 6 am), stores results in MongoDB, and publishes Allure reports to each engineer's network drive and to Azure DevOps, so the team can see exactly what failed, rerun individual tests, and filter by product area.


Sep 2026

TODO: a sample of generated text and a loss-curve image

GPT from scratch

A GPT-style decoder-only transformer implemented from scratch in PyTorch, following Andrej Karpathy's Let's build GPT lecture (token and positional embeddings, multi-head causal self-attention, pre-norm residual blocks, feed-forward layers with dropout) without high-level transformer libraries.

  • Trained character-level with AdamW, tracking train/validation loss to diagnose overfitting; scaled from a bigram baseline to a multi-layer transformer.

I wanted to actually understand how transformers work, the algorithm underneath every LLM, instead of just calling everything "AI" or "GPT."

TODO: GitHub repo URL


Feb 2023 – May 2026

Student Manager, MSOE IT Help Desk

TODO: a screenshot or short clip of the flow running on dummy asset tags

Repair workflow automation

Built a Power Automate + SharePoint pipeline that takes laptop asset tags, pulls device metadata from a central tracker, and drafts vendor repair emails, saving 4 hours of manual work each week. Managed scheduling, training, and device QA for 20+ student technicians.

Automating vendor repair requests saved four hours of manual work every week. It also gave me a reason to learn a new stack, and I put the time it freed up into my research.