Projects

Latent Diffusion Models for MNIST

Built an unconditional latent diffusion model on MNIST using:

• Autoencoders + Channel Attention Blocks
• U‑Net DDPM for denoising in latent space

Rock–Paper–Scissors Game Automation with YOLOv11

Real‑time hand gesture detection to power a fully automated Rock–Paper–Scissors game— complete with cheating detection and dynamic winner celebration.

• YOLOv11 object detection for Rock, Paper, Scissors gestures at ~98.2% mAP@0.5
• Cheating detection: red mask overlay & –1 point penalty for mid‑countdown changes
• Winner recognition: golden crown overlay with animated highlight

VAE for MNIST: Exploring Latent Spaces

Trained a Variational Autoencoder to compare reconstructions across different latent dimensions, illustrating the trade‑off between compression and fidelity.

Denoising Facial Emotion Dataset with Attention U‑Net and GAN

Addressed Gaussian and salt‑and‑pepper noise using an Attention U‑Net + PatchGAN pipeline, achieving high PSNR/SSIM for clearer emotion‑labeled images.

Reinforcement Learning Practices

Explored DQN, SARSA, D3QN, etc., with Boltzmann exploration optimizations. A hands‑on study of classic RL algorithms.

Fashion Tagger: AI‑Powered Fashion Image Labeling

Developed a multi‑label classification model trained on 44,000 fashion images, enabling real‑time tagging of fashion items. This tool assists in organizing and searching large fashion datasets efficiently.

Iranian Celebrity Face Recognition

Built a CNN‑based face recognition app (Flask + JS) to accurately identify Iranian celebrities, showcasing deep learning’s potential in real‑time image apps.

Automated README Generator with Multi‑Agent CrewAI

Utilized LLaMA 3 and CrewAI to automate the generation of comprehensive, stylistically consistent README files, streamlining documentation across repos.

Machine Learning Practices

Developed ML models for tasks like KNN, SVC, and Decision Trees, providing a hands‑on survey of standard supervised learning techniques on diverse datasets.