Gemma from Scratch

A minimal PyTorch implementation for training a Gemma-like language model from scratch. It supports inference with official Gemma weights and training a model on custom data. The model architecture is a decoder-only transformer with features like SwiGLU activation, RMSNorm, and Rotary Positional Embeddings (RoPE).

Machine Learning Highlights: Language Model, PyTorch, Transformer, Gemma

Technology Highlights: Python, PyTorch

Deep Learning for Tabular Data

An updated (2025) guide to Deep Learning for tabular data, comparing a fine-tuned Keras 3 (PyTorch backend) DNN and an Optuna-optimized XGBoost model on the Kaggle Amazon Employee Access Challenge. The project features a modernized tech stack, corrected cross-validation, hyperparameter optimization with Optuna, and an advanced DNN architecture.

Machine Learning Highlights: Deep Learning, Tabular Data, Keras, PyTorch, XGBoost, Optuna

Technology Highlights: Python, Keras, PyTorch, XGBoost, Optuna

Lunar Lander 2025

Revisiting an old project thanks to vibe coding and Gemini CLI. This project implements a Deep Q-Network (DQN) agent using Keras 3 with the JAX backend to solve the Lunar Lander environment from the Gymnasium library. The goal is to successfully land a spacecraft on a moon landing pad.

Machine Learning Highlights: Deep Q-Network (DQN), Reinforcement Learning, Keras, JAX

Technology Highlights: Python, Keras, JAX, Gymnasium

Gemini Hallucination Detection

Demonstrating how to use Gemini to detect hallucinations in LLMs. This project explores methods for comparing LLM-generated answers with target answers using text embeddings to score and detect hallucinations. It utilizes the "GenAI Hallucinations" dataset from a Kaggle competition.

Machine Learning Highlights: Large Language Models (LLMs), Gemini, Hallucination Detection, Text Embeddings

Technology Highlights: Python, Gemini

Naive Bayes from Scratch

How to create a gaussian naive bayes algorithm from scratch in Python and C++. The primary application demonstrated is predicting vehicle behavior on a highway (e.g., changing lanes or maintaining the current lane) using sensor data.

Machine Learning Highlights: Naive Bayes, Classification

Technology Highlights: Python, C++

Vibe Coding Experiment

W&B Founder and Vibe Coding Hands-on – W&B Meetup #22 in Tokyo. This project is a creative coding experiment focused on generating and exploring Julia fractals. It explains Julia fractals as escape-time fractals generated by repeatedly applying a simple mathematical formula in the complex plane.

Machine Learning Highlights: Fractals, Generative Art

Technology Highlights: Python