• Working on a time-series data of petroleum industry to detect various anomalies using different ML and statistical techniques such as SVC, Random Forest, PCA, Hotelling T-Square, Q-Statistic

• Created a POC for a Deep Learning based solution to the problem of uncalled No-Ball in cricket, We used YOLO - v3 to detect foot of the bowler and an Analytical algorithm to detect the landing point of bowler's foot.

• Responsible for analyzing sales data of top US drug manufacturers. Segmented prescribers using various clustering techniques based on prescribing patterns of doctors to maximize revenue.

• Built R - Shiny Applications that helped major pharma companies to better contract with insurance providers. Using these applications companies could make decisions within minutes which used to take weeks and save millions of dollars with the accuracy of predictions.

• Built Credit score models with 85% accuracy currently serving 100,000 customers to get instant loans within minutes.

• Designed wide-band microstrip patch antenna, (2 GHz to 10 GHz) using IE3D software.

• Compared different antenna designs on the basis of different parameters (gain, radiation pattern, beamwidth) and proposed two antenna designs with better characteristics.

• Fabricated both proposed antenna design

Read paper here

Abstract: Automatic License Plate Recognition (ALPR) has been a topic of research for many years now due to its real-life application but hasn’t been any significant breakthrough due to limitations in image processing algorithms to satisfy all the real-life scenarios such an illumination, moving cars, background etc. This paper presents a robust and efficient ALPR system using a combination of the ‘You only Look Once’ (YOLO) neural network architecture and standard Convolutional Neural Network (CNN). In total 3 stages of YOLO and 1 stage of CNN has been used in the proposed system. The last stage of YOLO and CNN have been specifically designed to perform detection (segmentation) and recognition of characters, respectively. We have built our own dataset of 604 car images in natural settings with different lighting conditions and viewing angles for the YOLO stages. In addition, a computer-generated dataset of 42237 characters has been used to train CNN. The resulting system has been tested on 50 random test images not part of training or validation datasets. The validation accuracies of all 4 stages exceed 90% whereas, the overall final accuracy on 50 test images comes to 82% with some fault tolerance. The use of deep learning instead of Image Processing also enabled to detect skewed license plates. The accuracy of stages 1 and 2 of YOLO were 100% on both validation and test sets.

Abstract: Object detection and localization is one of the prominent applications of the computer vision. The paper presents comparative study of state of the art deep learning methods - YOLOv2, YOLOv3 and Mask R-CNN, for detection of birds in the wild. Detection of birds is an important problem across multiple applications including the aviation safety, avian protection and ecological science of migrant bird species. Deep learning based methods are very pre-eminent at detecting and localizing the birds in the image as it can tackle the conditions wherein the birds shown are diverse in shapes and sizes and most importantly the complex backgrounds they are in. We used the training and testing dataset provided by the NCVPRIG (BROID) conference which contained 325 and 275 images respectively. For training, we used the pre-trained models on the VOC 2012 and COCO dataset and trained them on the 325 images. We used F – score as one of the performance metrics, and F-Scores were 0.8140, 0.8721, 0.8688 for the YOLOv2, YOLOv3 and Mask R- CNN respectively. The results show that YOLOv3 out performs YOLOv2 and is a marginal improvement over Mask R-CNN.