Md Ashfaq Salehin

I developed a novel approach that combines temporal graph neural networks (TGN) with transformer-based forecasting for microservice workload prediction. Leveraging Alibaba's 2022 microservice dataset, I achieved improved accuracy compared to existing methods, especially under peak traffic conditions. I designed temporal embeddings to capture dynamic service interactions, demonstrating enhanced long-term prediction capabilities.

I developed Temporal Walk, a high-performance library for constructing and sampling temporal networks to enable efficient training of graph neural networks (GNNs). The library is designed for scenarios where graph structures evolve over time, such as social networks, communication systems, and financial transaction graphs.

My key contributions include implementing efficient algorithms in C++ for temporal walk sampling and incorporating various biasing strategies such as uniform, linear, and exponential. I supported dynamic graph pruning with a sliding time window to optimize memory usage in long-running systems. Additionally, I exposed functionality through Python bindings using PyBind11, enabling seamless integration into data science workflows. I also focused on optimizing scalability, ensuring the system could process large-scale temporal graphs in real time.

This project plays a critical role in applications like dynamic node embedding, anomaly detection, and temporal link prediction. My responsibilities encompassed the full development lifecycle, including architecture design, implementation, performance tuning, and API integration.

I developed a transformer-based learning-to-rank (LTR) model to optimize listing ranking, focusing on price adjustments to improve visibility and user engagement.

My key responsibilities included preprocessing large-scale scraped listing rank data and implementing an LTR pipeline in PyTorch using pairwise hinge loss and ranking accuracy. I evaluated model performance using pairwise hinge loss and MRR metrics, ensuring a comprehensive assessment of the model's effectiveness. Additionally, I focused on optimizing the model for scalable, low-latency deployment to ensure efficient performance in production.

The project successfully improved listing ranking and click-through rates, showcasing the effectiveness of deep LTR models in large-scale platforms.

I developed an advanced reinforcement learning agent using Dueling Double DQN and neuroplasticity-inspired models. By integrating Hebbian plasticity, I enabled lifelong learning and compared the agent's performance against traditional methods. I advanced the field of adaptive systems by implementing plastic neural networks and testing them in the challenging domain of Atari games.

This personal project is an Android application for recording, editing, and mixing audio files. A limited open-source version of this application is available here: Github.com/ashfaq1701/fast-mixer.

Project Highlights:
• This is an NDK-based project. Most of the operational parts of the project are done as native codes.
• This project is built with Android architecture components, MVVM, LiveData, navigation, data binding, coroutines, and more.
• Hilt is used as a DI framework.
• RxJava and a reactive programming pattern are used to act against various view events.
• Used Google Oboe as the audio streaming framework and FFmpeg for decoding and encoding audio.

I developed this application as part of my job responsibilities while working at OPN. I built many core features of the application with native Kotlin MVVM architecture. and, as part of my job role, I owned some important features of this application.