Resume

PhD in Electrical & Computer Engineering

• Research Areas - Data Efficiency, Efficient Fine-tuning, Federated Learning, Large Language Models
• Research Advisor - Prof. Kaushik Roy

MS in Electrical & Computer Engineering

• Thesis - Energy Efficient Byzantine Agreement Protocols for Cyber-Physical Resilience

Bachelor of Engineering in Electronics and Communications

TRIM - Token-wise Attention-Derived Saliency for Data-Efficient Instruction Tuning

• Proposed a forward-only, gradient-free data selection method that builds attention-based token “fingerprints” to score instruction-tuning examples.
• Selected instruction examples that most improved performance on reasoning and commonsense benchmarks for open-source LLMs under small data budgets.
• TRIM-selected coresets outperformed state-of-the-art coreset baselines by up to ~9% on downstream tasks and, in some settings, matched or surpassed full-data fine-tuning.
• Demonstrated computational scalability- in a shared coreset selection setup, TRIM achieved higher downstream accuracy while running ~2.6× faster than the best gradient-based baseline.
• Under Review

Coresets from Trajectories - Selecting Data via Correlation of Loss Differences

• Introduced Correlation of Loss Differences (CLD), a gradient-free coreset metric that ranks training points by how closely their loss-change trajectories correlate with a small validation set.
• Established a convergence guarantee- training on CLD-selected coresets tracks full-data optimization up to a provable error bound.
• On CIFAR-100 and ImageNet-1k, achieved state-of-the-art coreset selection performance, typically within ~1% of more computationally expensive methods across subset sizes.
• Demonstrated cross-architecture transfer- coresets selected with small proxy CNNs generalized to larger CNN and vision-transformer models with <1% accuracy drop.
• Accepted for publication at TMLR 2025.

TOFU - Federated Learning with Data and Communication Efficiency

• Proposed a federated learning scheme that encodes each client's model update as gradients on a small synthetic dataset, transmitting proxies instead of full weight updates.
• On MNIST and CIFAR-10, achieved up to ~4x and ~6.6x lower communication cost than a standard federated learning baseline, while maintaining comparable final accuracy.
• Enhanced privacy against gradient inversion attacks- proxy data and reconstructed inputs resemble noise and fail to reveal meaningful client information.
• Studied accuracy-communication-privacy trade-offs by varying proxy size and update frequency, yielding practical guidelines for bandwidth-constrained federated deployments.
• Published in IEEE Access 2024.

DOTIE - Energy-Efficient Object Detection Using Event Cameras

• Developed an event-camera object detection pipeline using a lightweight spiking layer plus density-based clustering to isolate moving objects without frame reconstruction.
• On the MVSEC outdoor driving dataset, more than doubled the mean IoU over prior event-based methods and achieved near-perfect foreground detection.
• Achieved roughly six orders of magnitude lower energy consumption and significantly reduced latency compared to a YOLO CNN baseline on the same data.
• Maintained performance across diverse scenes with minimal retuning, enabling deployment in low-power autonomous navigation and neuromorphic systems.
• Presented at ICRA 2023 and CVPR 2023 Workshops.

Research Intern, Integrated Systems Team

• Prototyped an unsupervised anomaly detection pipeline for automated target recognition using state-of-the-art methods via Anomalib on MVTEC-AD and internal sensor datasets.
• Built an interactive labeling tool combining classical computer vision and SAM to generate pixel-wise masks on noisy imagery, reducing manual annotation time.
• Exported top models to ONNX and used the Latent AI Efficient Inference Platform (LEIP) to compile and optimize them for edge devices, achieving 4x energy-efficiency gains over unoptimized baselines.
• Summarized experiments and findings in internal documentation and presentations to support ongoing edge inference pipeline development.