Software Engineer · AI Systems

Building AI technologies with AI tools.

I learned Computer Science first, then learned the inside of LLMs by hand. Now I build with both: the substrate underneath and the tools on top.

To understand how Transformers and GPTs actually work, I hand-coded them — following Karpathy's Neural Networks: Zero to Hero end-to-end — before I let AI tools touch my workflow at all. Once I understood the substrate, I started using AI for development gradually and deliberately.

To work the way real teams do, I held to professional practices — feature branches, pull requests that carry the reasoning, versioned releases, decisions documented as I made them.

That sequence matters. When an AI tool is wrong, I notice. When a model fails silently, I know how to diagnose it. The work below is built on that footing.

Selected work

Where models meet real engineering.

Engineering case studies. Each project links to a full technical write-up.

01

Rocket Engine Health Orchestrator

A rocket engine usually fails quietly first — vibration creeping up, a temperature trending warm, a pressure easing down, each still inside its own limit. Conventional redline monitoring is structurally blind to that. So I'm building an advisory layer that reads the channels together and flags the combined signature before any one of them breaks.

PythonPyTorchTransformersCNNThree.js

2026 · in active development

Interactive demo
02

Turbofan Predictive Maintenance

I wanted real experience with what Transformers do beyond LLMs — so I picked a NASA problem where failure means someone dies: predicting a jet engine's remaining useful life.

PythonPyTorchTransformersFlask

2025–2026 · in active development

03

AskMickey

I grew up going to Disney World and figured there had to be an easy, fun way to get any park info you wanted. Building it right meant putting a deterministic routing layer in front of the LLM — not just calling Gemini and hoping.

DartFlutterGemini

2025 · iOS · Android · Web · Desktop

04

Weather Forecasting: Physics vs. ML

I've always loved meteorology — so I built a side-by-side comparison to put the claims to the test: physics-based GFS versus ML models from ECMWF AIFS and DeepMind WeatherNext, with a three-model accuracy scoreboard verified against real observations.

JavaScriptPythonNWSOpen-MeteoBigQueryECMWF AIFSDeepMind WeatherNext

2026 · live demo

05

Pong from Pixels

A deep Q-network that learns Pong from nothing but the screen — no game state, just raw pixels — trained from scratch on a 6 GB consumer GPU. I'm hand-building the network, replay buffer, and training loop, and shipping the honest learning curve: the long flat stretch, the silent bugs, and the climb when it finally starts to see.

PythonPyTorchGymnasiumCUDA

2026 · in active development

Toolkit

The whole stack, not just the top of it.

Foundations

  • Neural Networks
  • LLM Internals
  • Transformer architecture
  • Attention mechanisms
  • Tokenization · BPE
  • Backpropagation
  • Karpathy — Zero to Hero

AI / ML

  • PyTorch · Transformers
  • Generative AI · LLMs
  • RAG · Agentic AI
  • Multi-Task Learning
  • Prompt Engineering
  • Structured Outputs
  • Scikit-learn

Languages

  • Python
  • C · C++
  • JavaScript
  • Dart
  • Java · Swift
  • SQL
  • HTML · CSS

Frameworks & Tools

  • Flutter · Flask
  • Pandas · NumPy
  • Matplotlib
  • Git
  • Linux · UNIX
  • Amazon EC2
  • Arduino · Raspberry Pi

Background

William Opyrchal

Experience

ML Software Engineer Intern

Epcom Corporation · Summers 2024–2025

Researched and built a Transformer-based predictive maintenance system for multivariate time-series sensor data. Integrated retrieval-augmented LLMs with maintenance records and failure-mode data using prompt engineering and structured outputs; built data pipelines and evaluation harnesses for iterative model development and benchmarking.

Software Development Intern

Epcom Corporation · Summers 2020–2023

Built full-stack components of an Applicant Tracking System — Linux · Apache · MySQL · PHP · JavaScript · Bootstrap, deployed on Amazon EC2. Wrote automated cron jobs with pattern recognition to ingest applicants from IMAP email.

Education

University of Florida

B.S. Computer Science · August 2025

St. Petersburg College

A.A., Cum Laude · GPA 3.7

Independent Coursework

Harvard CS50; MIT OpenCourseWare — Calculus, Differential Equations, Linear Algebra, Physics I–III, Artificial Intelligence (6.034); Stanford CS229 (Machine Learning).

Ready to Build.

Résumé View · Download