Garry Yuan

Software 1

NanoGPT PyTorch Transformers Self-attention Colab notebook
NanoGPT Tiny Shakespeare model
Notebook project

NanoGPT Transformer Notebook

A hands-on PyTorch notebook that rebuilds the core pieces of a character-level GPT model on Tiny Shakespeare, from tokenization and batching through masked self-attention and generation.

What

  • Loads and explores the Tiny Shakespeare dataset for language-model training.
  • Builds character-level tokenization with encode/decode mappings.
  • Walks through autoregressive prediction targets and context windows.

How

  • Uses PyTorch tensors, batches, and train/validation splits.
  • Implements causal masking so tokens cannot attend to future positions.
  • Builds self-attention and multi-head attention components step by step.

Result

  • Shows practical understanding of transformer internals beyond API usage.
  • Connects math concepts like scaled dot-product attention to runnable code.
  • Strong focused artifact for deep-learning and LLM fundamentals.
Open notebook GitHub

Software 2

Browser games JavaScript Three.js Mistral AI chat PWA manifest
CoolAIGame logo AI game platform

CoolAIGame

A browser-native game platform with a searchable game library, instant game launcher, progression stats, leaderboards, and an AI assistant experience.

What

  • Game discovery and launch flow for zero-download browser play.
  • Progression surfaces for level, XP, coins, and play streaks.
  • Dedicated game pages with screenshots, controls, settings, and sharing.

How

  • Static HTML/CSS/JavaScript front end with reusable registry-driven game data.
  • Three.js powered animated landing canvas.
  • Mistral-backed chat endpoint for the in-page assistant.

Result

  • Installable PWA structure with manifest and service worker.
  • Responsive library, search, filters, and recent-play sections.
  • Playable iframe-based game shell with fullscreen and related-game flows.
Open project Open game page GitHub

Software 3

Computer vision Python FastAPI Streamlit OpenAI analysis
Emotion AI Vision + live analysis
GitHub project

Emoticon Emergent

An AI emotion-analysis workspace with live camera, screen recorder, body-language, stress, lie-detection, and OpenAI-assisted interpretation modules.

What

  • Realtime emotion and body-language analysis from camera or screen sources.
  • Multiple focused analyzers for stress, interview recordings, and vision workflows.
  • Login, payment, and database modules around the analysis tools.

How

  • Python app modules split across FastAPI, Streamlit-style pages, and backend services.
  • Dedicated live camera, video, screen recorder, and AI vision analyzer scripts.
  • OpenAI and ChatGPT interface layers for richer interpretation.

Result

  • Large end-to-end prototype for multimodal AI analysis.
  • Clear separation between capture, analysis, authentication, and UI paths.
  • Strong showcase of practical computer-vision product thinking.
GitHub

Software 4

FastAPI PostgreSQL Redis Celery AWS S3
xFood Community baking API
Backend platform

xFood Community Baking Platform

A full backend API for a baking community platform, covering recipes, user profiles, image uploads, social posts, reviews, circles, comments, likes, and messaging.

What

  • Recipe sharing, baking posts, community circles, comments, likes, and reviews.
  • JWT authentication with user profiles, roles, and direct messaging.
  • Image upload pipeline for recipe and baking media.

How

  • FastAPI service with SQLAlchemy models and Pydantic schemas.
  • PostgreSQL persistence, Redis caching, and Celery background jobs.
  • S3-backed file storage with image optimization and resizing.

Result

  • Production-shaped API architecture with clear endpoint boundaries.
  • Docker and Railway deployment configuration included.
  • API surface broad enough to support a real social food product.
GitHub

Software 5

React TypeScript Vite Supabase Tailwind CSS
Prompt CAD Design tool prototype
Web app

Prompt Craft CAD

A Vite and React web application prototype for a prompt-driven CAD/design workflow, built with a modern TypeScript frontend stack and Supabase project structure.

What

  • Prompt-oriented interface for shaping design/CAD ideas through a web app.
  • Modern component system using shadcn-ui conventions.
  • Frontend-first product prototype suitable for fast iteration.

How

  • React and TypeScript app running on Vite.
  • Tailwind CSS styling with component configuration.
  • Supabase folder included for backend/data integration.

Result

  • Good example of product prototyping with a modern frontend toolchain.
  • Readable project structure with `src`, `public`, and Supabase boundaries.
  • Shows comfort moving from idea prompt to interactive interface.
GitHub

Software 6

Machine learning PyTorch NumPy EDA Transformers
ML Lab From scratch to PyTorch
Learning repo

Machine Learning Project Lab

A hands-on machine-learning repository that moves from beginner classifiers and regression into neural networks, image classification, embeddings, and a tiny transformer language model.

What

  • Titanic survival, house-price regression, spam detection, and EDA notebooks.
  • MNIST and CIFAR-10 image classification experiments.
  • Embeddings text classifier and tiny transformer language model.

How

  • NumPy implementations for linear and logistic regression from scratch.
  • PyTorch training loops for neural networks and CNNs.
  • Runnable project folders with shared requirements and smoke-test commands.

Result

  • Broad ML foundation across tabular, text, and image data.
  • Clear progression from fundamentals to deeper architectures.
  • Useful proof of independent learning and rebuild-from-first-principles practice.
GitHub

Software 7

Football analytics Feature engineering Leakage-safe ML Python package
Soccer ML Outcome prediction
Sports analytics

Soccer Match Predictor

A machine-learning starter system for predicting home win, draw, or away win using rolling team form, match stats, market odds, and modern football metrics.

What

  • Predict match outcomes using historical results and optional xG-style features.
  • Support home, draw, and away classes for upcoming fixtures.
  • Document practical free and paid football data sources.

How

  • Leakage-safe feature generation that only uses matches before each fixture date.
  • Scripts for fetching Football-Data.co.uk baselines, training, and prediction.
  • Package structure with `src`, `scripts`, examples, and docs.

Result

  • Research-backed structure for sports prediction experiments.
  • Extensible schema for xG, xA, PPDA, shots, possession, and odds.
  • Good example of careful ML evaluation thinking.
GitHub

Software 8

PyTorch MOSFET I-V data SPICE modelling Robustness tests
ANN MOSFET EE + ML crossover
Surrogate model

ANN for MOSFET SPICE Model Extraction

A PyTorch neural-network surrogate that learns MOSFET electrical behaviour from simulated I-V sweeps as a first step toward SPICE model extraction.

What

  • Generate synthetic MOSFET drain-current sweeps with device variation.
  • Predict drain current from bias and geometry inputs.
  • Explore model robustness under noisy measurement-like inputs.

How

  • Compact Shichman-Hodges style data generator.
  • PyTorch ANN trained on `log10(|Id| + eps)` targets.
  • Scripts for data generation, training, prediction, and robustness evaluation.

Result

  • Strong bridge between electrical engineering and applied ML.
  • Testable package layout with docs and pipeline tests.
  • Portfolio-ready example of simulation-to-model experimentation.
GitHub

BE(Hons) Electrical Engineering / Computer Science, UNSW, 2022-2026.