AI Research Portfolio

Hands-on, from-scratch implementations of modern deep learning architectures, from Vision Transformers to World Models and Joint-Embedding Predictive Architectures.

Focus: Self-supervised learning, world models, joint-embedding predictive architectures (JEPA), multimodal LLMs, and agentic systems.

Highlighted Experiments

Experiment	Paper	Key Result	Code
World Models (VMC)	Ha & Schmidhuber 2018	CarRacing reward: --	code
ViT from Scratch	Dosovitskiy et al. 2020	CIFAR-10 acc: --%	code
YOLOv1	Redmon et al. 2016	--	code

Results are filled in as each experiment completes training. Each experiment follows a standardized write-up format.

Topics

Topic	Description	Status
World Models	VMC, DreamerV3, DIAMOND	Active
Vision Transformers	ViT, MAE	Active
Computer Vision	YOLO evolution, detection	Active
Self-Supervised Learning	VICReg, SimCLR, Barlow Twins	Active
JEPA	I-JEPA, V-JEPA, MC-JEPA	Active
MLLM	CLIP, LLaVA — vision-language alignment	Planned
Reinforcement Learning	PPO, SAC, JEPA+RL	Planned
Agent + Tool Use	ReAct agent, Anthropic tool use, JEPA vision input	Planned
Embodied AI	Isaac Gym, JEPA Navigator	Planned

Technical Stack

• PyTorch — from-scratch implementations, no high-level wrappers
• uv + ruff + pytest — modern Python tooling (fast dependency management, linting, testing)
• Type hints and Google-style docstrings throughout

Repository Structure

ai-research-notes/
├── topics/
│   ├── world-models/              # VAE + MDN-RNN + Controller
│   ├── vision-transformers/       # ViT, MAE from scratch
│   ├── computer-vision/           # YOLOv1 model + loss
│   ├── self-supervised-learning/  # VICReg, SimCLR, Barlow Twins
│   ├── jepa/                      # I-JEPA, V-JEPA, MC-JEPA
│   ├── mllm/                      # CLIP, LLaVA (planned)
│   ├── reinforcement-learning/    # PPO, SAC (planned)
│   ├── agent-tool-use/            # ReAct agent, Anthropic tool use (planned)
│   └── embodied-ai/              # Isaac Gym, Habitat (planned)
├── docs/                          # Templates and guides
├── pyproject.toml                 # Shared dependencies
└── Makefile                       # Top-level commands

Getting Started

# Setup environment (installs uv if needed + syncs dependencies)
make setup

# Run tests
make test

# Run linter
make lint

# Train an experiment (example: VMC VAE)
uv run python -m topics.world_models.experiments.01_car_racing.src.train_vae