B.Tech ECE, VIT Vellore · Research Assistant @ MBZUAI · Incoming M.S.
I work on vision-language models for geospatial understanding — zero-shot instance segmentation of satellite and aerial imagery. Broader interests include computer vision, multimodal AI, NLP, and edge deployment.
I recently completed my B.Tech in Electronics and Communication Engineering from Vellore Institute of Technology (GPA: 8.29) and am preparing to begin graduate studies. I am currently a Research Assistant at the Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), where I am building Tera-SAM — an automated agentic segmentation pipeline leveraging SAM3 and vision-language models for zero-shot instance segmentation across heterogeneous geospatial sensor modalities including Gaofen-7 and SpaceNet.
Previously I have interned at IIT Guwahati (image restoration for weather-degraded imagery), the National University of Singapore (IoT anomaly detection with ensemble ML), IIT Dharwad (edge deployment on NVIDIA Jetson Nano), and VIT (edge computer vision systems for real-time surveillance). I enjoy working at the intersection of research and systems — building models that are both principled and deployable on real hardware.
I am open to research collaborations and always glad to discuss multimodal AI, remote sensing, or edge ML. Feel free to reach out.
Architecting Tera-SAM — an automated agentic pipeline using SAM3 and VLMs for zero-shot instance segmentation across heterogeneous sensor modalities (Gaofen-7, SpaceNet). Employs geometric and natural language prompt engineering; built end-to-end raster-to-vector alignment workflows with rasterio and geopandas.
▸ High-fidelity building mask generation without fine-tuning; superior zero-shot transferability across multi-spectral datasets demonstrated via mIoU metrics.
Image restoration for weather-degraded images, implementing a novel method to enhance quality while optimizing computational efficiency.
▸ Improved PSNR and SSIM by 5% over baseline while reducing resource requirements.
Advanced anomaly detection using ML and ensemble techniques to identify security threats in IoT devices.
▸ 0.952 accuracy in threat classification; 15% higher detection rate and 50% lower inference time vs. deep learning baselines.
Object detection and distance tracking on NVIDIA Jetson Nano with YOLOv8. Integrated flash-attn for vision-language model experimentation on edge hardware.
Edge CV system for reliable object detection in daylight and low-illumination environments.
▸ Real-time detection under varying lighting, identifying key objects within 0.5 seconds for surveillance and autonomous applications.
ML methods for high-cardinality categorical features in actuarial applications; supported data analysis and model development aligned with a published paper.
Agentic segmentation pipeline for satellite and aerial imagery using SAM3 as the mask generator and a VLM agent to iteratively refine prompts (points/boxes) based on segmentation errors. Includes automatic hard-case discovery and refinement loops for thin structures (roads), low-contrast regions, and cluttered scenes. Evaluated using IoU/Dice and qualitative failure-case analysis.
Raspberry Pi 5–based edge perception module using YOLOv5–v8 for nighttime vehicle detection, ranging, and driver alerts. Measures distances via Vision Eye and preemptively alerts drivers or adjusts autonomous vehicle trajectories.
Real-time intrusion detection and prevention for IoT networks using ML and ensemble learning. Achieves 0.95 accuracy and reduces inference time by 50% compared to deep learning baselines.
Real-time system using YOLO for object detection, eye-vision tracking for distance estimation, and a fine-tuned Florence-2 model on NVIDIA Jetson Orin Nano for low-latency audio descriptions of surroundings.
I am happy to discuss research, potential collaborations, or just chat about vision-language models, geospatial AI, or edge deployment. The best way to reach me is by email.