Enthusiastic Software Engineering student at NUST with a passion for Artificial Intelligence, and keen interest in Physics and Cognitive Science. Proficient in Machine Learning, Deep Learning, Computer Vision, and NLP. Eager to apply AI advancements to revolutionize agriculture, astronomy, healthcare, and education sectors, aiming to make a meaningful impact on society.
- 🖥️ See my portfolio at Here
- ✉️ You can contact me at [email protected]
- 🤝 I'm open to collaborating on Research and AI related projects including Machine & Deep learning, Computer vision, NLP, Generative and Agentic AI.
- Proposed MHA-UNET, an advanced architecture for CT angiography.
- Performance: Outperformed nnUNET and Swin Transformer in segmentation tasks.
- Expanding to multimodal inputs (CT images + relevant tabular and textual data).
- Performance: Outperformed nnUNET and Swin Transformer in segmentation tasks.
- Proposed ExCNET, a compact and efficient architecture for image:
- Integrated Squeeze-and-Excitation (SE) blocks for better feature extraction.
- Used Batch Normalization in between to improve model stability and training.
- Beating complex architectures like Deep-CNN , ResNet-50, ViT while maintaining a smaller model size of 8M.
- Evaluated open-source LLMs for Text-to-SQL tasks.
- Observed Llama 3.1 70B outperforming Mistral LLMs in generating complex queries.
- When using Llama 3.1 405B as a judge, system performance was further enhanced.
- Researching reasoning capabilities in Multimodal LLMs on medical datasets.
- Demonstrated strong reasoning and in-context learning performance on pneumonia X-ray images.
- Achieved better-than-expected results by exploring Chain-of-Thought (CoT) techniques.
- Investigating other techniques like Tree-of-Thought (ToT), etc to enable self-correction in multimodal models.
- Studying pretrained LLMs on tabular Data.
- Achieved better-than-expected performance with emergent behavior in predictions.
- While LLMs show promise, statistical models still outperform in certain structured data tasks.
- Investigated the role of Reinforcement Learning (RL) in improving LLM reasoning capabilities:
- Exploring RL-based feedback mechanisms for post-training optimization and its effectiveness for SFT (Supervised Fine-Tuning) data gathering and how scaling RL methods helps to achieve more robust performance improvements.
- Enhancing the reasoning capabilities of LLMs, VLMS, and multimodal systems by further exploring Reasoning techniques and exploring how effectively Reinforcement Learning can help in achieveing goals.
My GitHub Stats