← Back/Projects / Software and AI
Early AI and Machine Learning Work
Self-taught, high school, Python, PyTorch, scikit-learn, Fall 2023 to spring 2024
ProblemI wanted to know how the models in the headlines actually worked, not read about them.
My roleBuilt each one myself in Python, following tutorials and then changing things until they broke and I understood why.
ResultFour working projects on GitHub and a working knowledge of training loops, replay memory, feature extraction, and search.
LinksGitHub, GMebane525
| Role | Self-taught, high school |
|---|---|
| Context | Python, PyTorch, scikit-learn |
| Dates | Fall 2023 to spring 2024 |
| Field | Machine Learning, Computer Vision |
| Skills | Python, PyTorch, scikit-learn, NLTK, Pathfinding, Computer Vision, Reinforcement Learning, Java |
What I built
- Snake, deep Q-learning. A PyTorch agent that learns to play Snake in Pygame. Linear Q-network, replay memory, epsilon-greedy exploration. Follows a well-known tutorial closely, then I changed the reward shaping and network width to see what happened. A trained model is checked in
- A* pathfinding. Grid maze with eight-direction moves, open and closed sets drawn live in matplotlib so I could watch the search instead of trusting it
- Face recognition attendance. Webcam tracker that encodes reference photos with dlib, matches live frames, and logs names to a dated sheet. The most useful of the four and the one that taught me how brittle a vision pipeline is to lighting
- Sentiment analysis. VADER on Amazon reviews with a confusion matrix, then a TF-IDF plus logistic regression classifier on the NLTK movie review set
Honest scope
These are learning projects from senior year of high school. None is original research. They matter because they are where the AI tooling I now build for an investment firm and for ToleranceForge started, and because I can read a training loop or a model card and know what I am looking at.
The account also holds a year of AP Computer Science Java labs.