Skip to content

Repository files navigation

ML Adventures

An AI-powered learning project. I'm a senior software engineer (C++/Python, HPC, numerical simulation) building ML skills through hands-on projects — with an AI as teacher, not as substitute.

Highlights

  • 06 — Equation-to-Code (QLoRA on Qwen2-VL) — fine-tuned a Vision-Language Model (VLM) to convert images of typeset math equations into executable Python code. 91.0% validation accuracy.
  • 05 — Character-Level GPT — implemented a decoder-only transformer from scratch (attention, positional encoding, autoregressive generation) following Karpathy's nanoGPT walkthrough, trained on GPU. 2.24 train / 2.27 val loss.

Background

I have 8+ years of production C++/Python, a PhD in Computational Mechanics, and HPC/distributed computing experience. Strong applied math foundations (linear algebra, optimization, PDEs). I've done tutorial-scale and self-directed ML projects, but no professional ML experience yet. This repo is my path from numerical software engineer to ML engineer.

How this works

  • I write the code. The AI provides direction, asks questions, reviews my work, and pushes back when I'm wrong.
  • Framing, design, and task documents are written by the AI under my supervision.
  • Projects are real and complete — each one trains a model, shows results, and has a clear README.
  • Training runs on a local GPU (RTX 4070 Ti) or Kaggle/Colab free-tier GPUs.
  • AI steering lives in AGENTS.md — it sets ground rules for how the AI interacts with me (teacher, not substitute).

Progress

# Project Status Key Result
01 Image Classifier (CIFAR-10) 85% test accuracy
02 Tic-Tac-Toe (Tabular Q-Learning) Never loses vs random
03 Tic-Tac-Toe (DQN) Never loses vs random
04 CartPole (DQN) 500/500 perfect
05 Character-Level GPT 2.24 train / 2.27 val loss
06 Equation-to-Code (QLoRA Qwen2-VL) 91.0% validation accuracy
07 Physics-Informed AI (FEM + PINN, cantilever beam) 🚧 FEM solve validated vs. Euler-Bernoulli; PINN in progress

Roadmap

Phase 1: Core ML & RL

Get fluent in PyTorch, modern training workflows, and Reinforcement Learning (RL) fundamentals.

  • Image classifier (CNN on CIFAR-10) — 85% test accuracy
  • Tabular Q-learning agent (Tic-Tac-Toe — Bellman equation, self-play)
  • Deep Q-Network (DQN) agent (Tic-Tac-Toe — replay buffer, target network)
  • DQN on CartPole (single-agent, dense reward, Gymnasium environment)

Phase 2: Transformers & Embeddings

Understand attention, embeddings, and generative models.

  • Character-level GPT (build a transformer from scratch, train on GPU)
  • Fine-tune a small language or vision model on a custom task
  • Semantic search / retrieval system

Phase 3: Specialization

One or two deeper projects — pick based on interest and job market at the time.

  • Diffusion models (image or video generation)
  • World models (action-conditioned environment prediction)
  • ML infrastructure (distributed training, model serving, custom CUDA kernels)
  • Scientific ML (physics-informed neural networks, surrogate models)
  • RL at scale (Proximal Policy Optimization (PPO) / policy gradient on a visual environment)

About

ML quests: zero to hero. The AI is the dungeon master

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages