🚀 Exploring AI at Scale | Building Smarter Models | Advancing AI Research
I am a Machine Learning Researcher with a passion for AI, deep learning, and large-scale knowledge systems. Currently pursuing an MS-Research in Computer Science at IIIT Hyderabad, my work focuses on Geometric Deep Learning, Temporal Graph Learning, and Large Language Models (LLMs). I specialize in federated learning, adversarial robustness, and scalable AI architectures, aiming to bridge the gap between cutting-edge research and real-world AI applications.
With 3+ years of experience in AI and data science, I have worked at Barclays as a Macro Structuring Analyst, where I developed AI-driven predictive models for financial markets, optimizing risk assessment and trading strategies for 10,000+ trades across 25 strategies. My expertise in LLM fine-tuning, AI model optimization, and knowledge graphs extends into both research and industry applications.
I have worked on federated learning, fine-tuning large models, and optimizing LLMs for efficiency and security, with publications in top AI venues. My experience extends to applying AI in real-world settings, from financial modeling at Barclays to AI-driven systems for robust and scalable inference.
Introduces a novel framework for federated learning on graph-structured data, addressing challenges of heterophilic, non-IID graphs in decentralized settings. Our approach leverages spectral graph transformations for robust feature aggregation while preserving global graph properties, enabling efficient cross-client knowledge sharing without direct data exchange. The framework demonstrates significant improvements in accuracy and convergence stability compared to standard federated GNN approaches, with applications in healthcare networks, financial risk modeling, and decentralized knowledge graphs.
Introduces a novel adversarial attack framework for Temporal Graph Neural Networks (TGNNs), exposing vulnerabilities in dynamic graph-based AI systems. Unlike traditional attacks that focus on node or feature manipulation, TERA strategically targets temporal edge rankings, disrupting the structural and temporal dependencies crucial for model predictions.
By modifying edge importance over time, TERA significantly degrades model performance while maintaining stealth and minimal perturbation, making detection challenging. Our experiments across multiple real-world temporal graph datasets demonstrate a 20.3% drop in predictive accuracy across diverse TGNN architectures, highlighting critical weaknesses in time-sensitive AI applications. These findings emphasize the need for robust adversarial defenses in dynamic graph learning, especially in domains such as financial transaction monitoring, cybersecurity, and social network analysis.
Explores a novel abstention-based learning strategy to improve prediction reliability in Temporal Graph Neural Networks (TGNNs). Dynamic graphs often suffer from noisy, imbalanced, or uncertain data, leading to erroneous predictions in critical applications like fraud detection, social networks, and financial modeling. Our approach introduces a confidence-aware abstention mechanism, allowing the model to defer uncertain predictions, thereby enhancing overall robustness.
By integrating adaptive confidence thresholds and selective prediction strategies, our method ensures that the model prioritizes high-confidence predictions while avoiding misleading outputs. Experimental results on multiple temporal graph benchmarks show that this approach improves model reliability and decision-making, particularly in low-data or high-noise scenarios. This work highlights the potential of abstention as a tool for enhancing trustworthiness and accuracy in dynamic graph learning applications.
Combined Large Language Models (LLMs) with Graph Neural Networks (GNNs) to enhance reasoning over graph-structured data, introducing Subprocess and Island encodings. Subprocess encoding enabled directed graph reasoning, outperforming prior methods, while Island encoding improved global graph structure understanding by 25%. Conducted systematic evaluations across GPT-4, Claude, Llama, and Gemma LLMs, optimizing graph-to-text representation strategies for diverse reasoning tasks.
Leveraged Vision-Language Models (VLMs) as the backbone for zero-shot classification, extending Jiang et al.'s knowledge erasure algorithm to incorporate internal confidence scores. Our approach achieved robust classification performance on the CIFAR10 dataset by focusing on aggregating patch-level confidence scores. The research demonstrated improved robustness against adversarial settings, such as Gaussian noise and brightness variations, highlighting the potential of VLM-based confidence mechanisms.
💡 Machine Learning & AI: LLMs, NLP, Graph Neural Networks, Diffusion Models
🛠 Tech Stack: Python, PyTorch, TensorFlow, JAX, Hugging Face
🌍 Scalable AI Systems: Distributed ML, Federated Learning, MLOps
📊 Research & Development: Model Evaluation, Adversarial Robustness, Multimodal AI