Gaurav Singh

Designed and implemented a cattle activity tracking system.

• Automatically analyze behavioral patterns from videos, and improve decision-making processes.
• Following solution components: cattle detection, tracking, occlusion handling, a centralized server, and video stream were implemented.
• Dashboard for activity insights and alerts.

Farm geofencing.

• Detected farms and other land types using Mask R-CNN.
• Extracted the boundary of each detected mask and calculated the latitude and longitude of each pixel boundary.
• CUDA, C++ programming, and multithreading are used to optimize services, which have improved response time by 70%.

Pest, disease, and deficiency detection in crops.

• Create one classifier model and one detection model (Mask R-CNN) for each crop. Classifier model for healthy and anomaly classification, and detection of anomalies.
• Used TensorFlow Serving to serve the model, which has improved the response time of the service by 50%.

Personalized Advisory to Farmers.

• Provides personalized and context-aware agricultural advisory to farmers.
• Fine-tuned Llama2 model for chat response, context processing, and created RAG from the appropriate advisory database.

Attribute mapping of retail products.

• Fully fine-tuned/trained two LLM-based models, one for identifying the category of products, and the other to map attribute values into the format of AJIO.
• Realized cost savings of 62% in manual mapping processes.

Researched and implemented smart image cropping for product cataloging.

• Automatically identify and crop the most relevant portions of an image, enhancing visual focus, and improving optimized and efficient content for OCI cloud infrastructure.
• Used saliency detection, edges using Laplace, and regions high in saturation.

Innovated methods to streamline automatic document structure.

• Automatically organizes text documents into appropriate directories based on content and context.
• Created a transformer-based BERT model.

Enhanced searchability of products in an e-commerce-based application on OCI infrastructure.

• Enhancing visual content understanding, metadata generation, and better tagging of images for e-commerce products.
• Used image caption technique.

Research on object detection in point cloud/LiDAR data for autonomous vehicles.

• Developed cutting-edge solutions for identifying and classifying objects in three-dimensional environments.
• Used deep learning and linear algebra-based features from LiDAR data to create a new deep learning model.
• Tested other SOTA algorithms, like PointNet++ and SVM.

Road Sign Detection and Geo-Tagging Using Camera Calibration.

• To automatically detect road signs from images or video streams captured by cameras installed on vehicles or traffic infrastructure.
• This technology has significant implications for road safety, traffic management, and navigation systems.

Lidar data visualization and classification tool.

• Engineered a C++ tool enabling the visualization and automatic classification of LiDAR data points.
• In tool development, C++ and its libraries, OpenGL, GDAL, etc., have been used.

Employee identity verification using facial recognition and ID numbers with existing records in the database.

• Implemented an advanced verification system combining facial recognition and ID number checks.
• Pre-indexing embeddings with approximate nearest neighbor (ANN) search is used for fast processing.
• Facenet and OCR have been used for face and ID number, respectively.

Implemented the Home Credit Risk Analysis system.

• Assessing the creditworthiness of customers and recommending credit positivity based on their credit records from different portfolios.
• Designed features including credit utilization ratio, days past due metric, income-to-debt ratio calculation, and portfolio diversity index evaluation.
• Implemented a classification and regression model to predict risk levels and scores of continuous credit risk levels, respectively.

Credit card transaction fraud detection.

• Developed system to detect fraudulent transactions instantly and accurately, reducing false positives.
• Train models using historical labeled data (fraudulent vs non-fraudulent).
• Applied XGBoost, LightGBM, Random Forest algorithms as well as CNN and RNN deep learning models for handling sequential data.

Parking lot detection for satellite images POC.

• Implemented a POC to identify parking lots, and counted the number of available parking spaces.
• Used OpenCV for thresholding, Canny edge detector, contour detection

Developed a rule-based TV channel suggestion system.

• To provide viewers with personalized channel package recommendations based on predefined rules and user preferences.
• Curated content tailored to past viewing history and genre preferences.

Statistical Analysis of Channel Demands.

• To understand the demand for products or services through different distribution channels.

Backend development of a web-based TV channel subscription application.

• Worked on backend development and used Java, Spring Boot.
• Special package subscription module.