About
I am an AI/ML Engineer with a focus on Deep Learning R&D. I graduated with a Bachelor's degree in Computer Science from FAST NUCES in 2025. My current work centers around Hyperspectral Imaging, where I explore the intersection of deep learning and spectral analysis. Over the years, I've worked on AI research, built intelligent systems, and experimented with cutting-edge models across domains such as Computer Vision, LLMs, and Generative AI. I am passionate about solving real-world problems through technology and continuously expanding my skill set in applied machine learning and research-driven development.
Skills
Things I’ve made trying to put my dent in the universe.
I’ve worked on tons of little projects over the years but these are the ones that I’m most proud of. Many of them are open-source, so if you see something that piques your interest, check out the code and contribute if you have ideas for how it can be improved.
Ad-Pilot
Created a marketing automation platform tailored for small businesses to manage ad campaigns across Facebook and Instagram.Ad-Pilot streamlines content creation, automated scheduling, real-time performance insights, and competitor analysis to enhance marketing efficiency.
WhatsWhisper
Created a multi-featured WhatsApp bot that transcribes voice messages into text using advanced models like OpenAI's Whisper and Alibaba's ZipEnhancer. Ideal for noisy environments or when listening isn't possible, while also offering smart task scheduling and leveraging Acoustic Noise Enhancement for improved transcription accuracy.
DiReCT-RAG: Diagnostic Reasoning in Clinical Notes
Designed and implemented a Retrieval-Augmented Generation (RAG) pipeline for clinical diagnostic reasoning. Utilized semantic chunking, dynamic confidence scoring, and advanced LLM integration (Palmyra-Med-70B-32k via NVIDIA AI Endpoints). Integrated Google Generative AI Embeddings and Clinical ModernBERT for dual retrieval support. Evaluated with Gemini 2.0 Flash (LLM-as-a-Judge) on MIMIC-IV-Ext clinical notes.
Finetuned DeepSeek R1 with GRPO for Emoji based Math Challenges
Built an interactive web application that solves mathematical problems written with emojis using a fine-tuned DeepSeek-R1 language model with GRPO (Group Relative Policy Optimization). The model specializes in solving creative mathematical equations represented by emojis, with an easy-to-use Streamlit interface.
Research Contributions & Academic Work
A collection of my research papers and academic contributions in the fields of Deep Learning, Hyperspectral Imaging, and Computer Vision. These publications represent my commitment to advancing the field through rigorous research and innovative solutions.
Differential Attention with Enhanced Squeeze-and-Excitation for Hyperspectral Image Classification
Saad Sohail, Muhammad Usama, Usman Ghous, Manuel Mazzara, Muhammad Ahmad
IEEE Geoscience and Remote Sensing Letters (IEEE GRSL) (2025)
This paper introduces DIFF-SE, a novel transformer-based framework for hyperspectral image classification (HSIC), combining multi-head differential attention and an Enhanced Squeeze-and-Excitation (E-SE) module. Differential attention contrasts attention maps to boost discriminative features while suppressing noise and redundancy. The E-SE module recalibrates spectral bands and spatial regions, enhancing critical features. DIFF-SE achieves up to 99.34% accuracy on benchmark datasets (Pavia University, WHU-Hi-HanChuan, OHID-1), showing strong generalization and robustness in complex, noisy environments.
EnergyFormer: Energy Attention with Fourier Embedding for Hyperspectral Image Classification
Saad Sohail, Muhammad Usama, Usman Ghous, Manuel Mazzara, Salvatore Distefano, Muhammad Ahmad
IEEE Geoscience and Remote Sensing Letters (IEEE GRSL) (2025)
This novel approach introduces Energy Attention, a mechanism designed to enhance feature learning by integrating energy spectral density into the attention process. Additionally, it leverages Fourier Embedding to capture spatial-spectral correlations, effectively improving classification performance.Why does this matter? Traditional attention mechanisms often struggle with the high-dimensional complexity of hyperspectral data. EnergyFormer overcomes this challenge by incorporating spectral energy information, making it more robust, efficient, and accurate.
Recent Updates & Announcements
Stay updated with my latest achievements, publications, project launches, and professional milestones.
June 2025
Accepted for Masters in Artificial Intelligence at LUMS
Got accepted into the Master's in Artificial Intelligence program at the Syed Babar Ali School of Science and Engineering (SBASSE), Lahore University of Management Sciences (LUMS).
June 2025
Paper Accepted at IEEE GRSL
My first paper 'Differential Attention with Enhanced Squeeze-and-Excitation for Hyperspectral Image Classification' got accepted for publication in IEEE Geoscience and Remote Sensing Letters.
Get in Touch
Want to chat? Just shoot me a dm with a direct question on my Email and I'll respond as soon as possible.