Murad Asgarov

Detail-oriented individual with a focus on effective time management and organizational strategies. Ready to leverage skills in various settings to support operational efficiency. Aiming to drive company success through dedicated performance.

• Implemented a self-supervised JEPA architecture in PyTorch, using a Vision Transformer to predict masked target embeddings from visible context patches. The model leverages EMA-updated target encoders and trains in latent space without pixel reconstruction, enabling strong semantic representation learning.

• Built a toy Transformer model in TensorFlow to simulate latent state reasoning by inserting additional "thought steps" conditioned on custom tokens during sequence generation. Implemented multi-step reasoning with optional token-based injection, allowing selective expansion of the hidden representation for logic-aware language modeling. This project explores interpretability and token-triggered inference in a lightweight, controlled setting.

• Developed a neural style transfer system in PyTorch using a pretrained VGG19 to extract multi-layer content and style features. The model optimizes a randomly initialized image to match the content of one image and the style (including color statistics) of another, with dynamic weighting of losses over time. Included total variation loss and custom scheduling of style/content weights for enhanced visual coherence.

• Implemented a custom InfiniAttention layer in TensorFlow, combining local dot-product attention with a learnable memory mechanism for long-term dependencies. The module adaptively blends classical attention with recursive updates using a sigmoid-weighted controller and supports delta-rule correction. Designed as a reusable component for future research and experimentation with sequence models beyond standard transformers.

• Built a cryptocurrency data pipeline using the CCXT library to automatically fetch and save full historical OHLCV data for all Binance USDT trading pairs. Collected daily data from 2015 onward and exported each symbol to separate CSV files for downstream analysis or model training. The script dynamically handles API limits and missing data while maintaining a modular structure for extensibility.