Shikhar Shukla

Successfully created a high throughput end to end live video processing pipeline for parking lots from scratch using Numpy and PyTorch, that processes up to 40 live cameras on a single Tesla T4 GPU. Identified and integrated ML models into NeuralMarker (Data Labeling Platform) to make data labeling process 150% faster. Developed an in-house model deployment pipeline. Trained a segmentation model in PyTorch to segment out tree branches and determine if a parking spot is visible from a camera. Trained and integrated the EasyOCR model into the high throughput pipeline to log all cars entering and exiting a parking lot. Used Superglue Model to detect and rectify camera movement in the field due to hailstorms and other extreme weather conditions. Trained a classification model to detect bad weather conditions in a parking lot and integrated and deployed it in production. Used Homography to Translate Camera stream into GPS coordinates. Built a Dataset Drift Detection System to help find out online Anomaly cases to aid model training. Built a Data Flywheel to collect and label data for all of the above-mentioned models using features from a ResNet18 backbone and k-means clustering. Experimented with 3D Object Detection to detect cars in 3 dimensions. Experimented with Unity Computer Vision and Unreal Engine to generate Synthetic Data for faster model training.

Tested Flutter for cross-platform app development. Developed a PhotoOCR app using Firebase ML kit.