Karan Pathak

FloodMetaSegNet - Flood Damage Extent Detection Using Satellite Images and Meta Attributes

• Co-developed FloodMetaSegNet (deep convolutional semantic segmentation model with meta channel) using PyTorch, GeoPandas, Rasterio and Shapely to determine the extent of building damage caused by floods using satellite images and elevation attributes.
• Applied transfer learning using renowned flood damage datasets to overcome the data shortage.
• Increased Dice score by 80% and Tversky score by 30% as compared with segmentation models without meta attributes.

Gardening E-Commerce Marketplace

• Collaborated with a 3-person team to develop a Gardening E-Commerce Marketplace solution using ReactJS, ExpressJS, PostgreSQL, and Prisma (ORM).
• Deployed the solution to AWS Fargate using AWS ECS service to handle a variable concurrent customer load.

Error logging module

• Designed and co-developed a NodeJS package to log information and errors from NodeJS cloud applications into different services like Axiom, GCP Cloud Logging, AWS CloudWatch, etc., which reduced the development time.
• Used in all the NodeJS cloud applications in the company.

Email Marketing Campaign Tool

• Designed a cost-efficient email marketing campaign system to manage customer leads and schedule email sending based on the customer’s persona without landing emails in the spam folder.
• Co-created microservices in ExpressJS to upload leads to the PostgreSQL database, manage email scheduling, and take actions based on customer responses.
• Increase B2B product sales by 38%.

WhatsApp Chatbot for Customer Engagement

• Designed a rule-based Shopify-WhatsApp chatbot to enhance customer engagement experience like impulsive buying, customer inquiry resolution, etc.
• Co-developed the application in ExpressJS and deployed it on Vercel (Serverless compute platform) and increased B2C product sales by 45%.

AI-Platform SDK

• Designed and created a Software Development Kit in Python to automate & manage the compute, data storage, and model deployment on the AI-Platform through the SDK.
• Helped platform's clients in running CPU & GPU-based data workloads and inference pipelines in development and production scenarios.
• Deployed and served AI models as API using BentoML

Chemical Raw Material Inventory Optimization using Time-Series forecasting

• Applied an LSTM model in PyTorch to forecast the consumption of individual chemicals in a complex series of chemical processes.
• Reduced the raw material inventory storage costs by 23% by achieving a mean MAPE score of 8.7%.

Alzheimer Research with Stall Catchers – Classify Stalled Blood Vessels In Brain Images

• Applied a CNN-LSTM network with a time-distributed layer on the Alzheimer video dataset to classify if a blood vessel is stalled or not in PyTorch.
• Achieved the best Mathews Correlation Coefficient of 0.548 with EfficientNet-B2 spatial encoder.

Fault geobody detection

Global model for fault geobody detection using deep supervised learning

• Built a supervised 2D deep CNN to detect fault geobody in seismic signals.
• Trained on multiple surveys with different distributions to achieve a global model.
• Used Unet as the model architecture.

Seismic Noise removal using V-Net

Built a supervised 3D CNN model to enhance fault detection by removing noise (small faults, false positives and non-laterally pervasive faults) from global fault model predictions.

Salt body detection using network-based deep transfer learning

• Built a model to detect the Top of Salt geobody using very few training data.
• Used Tiramisu architecture for semantic segmentation.
• Performed transfer learning by training only the last few layers of the decoder of a model pre-trained on seismic fault data and froze the remaining layers of the model.

Sea-Floor detection using weakly supervised deep learning

• Designed and developed a unique process for accurately detecting seafloor.
• Built a global model using weakly supervised learning to predict seafloor and applied a series of smart post processors on top of the model’s prediction to increase the overall accuracy.
• Filed provisional patent. U.S. patent no.: 62/914,608

Real-Time Hand Gesture-based Mouse Control

• Developed a novel rotation and scale-invariant pipeline for controlling mouse pointers using a webcam. Demo
• Defined custom hand gestures to imitate all the current mouse functionalities, including the drag-and-drop feature.

Client Library for Logstore

Logstore is a distributed logs storage system used to store well logs of an Oil Field. I worked on building client library for logtore, and integrated logstore with different microservices developed by my team using the client library.

DLIS parser

A DLIS file is a log file created by sensors that monitor petroleum wells. I worked on creating a DLIS parser for extracting log headers and data from a DLIS file and store the information extracted in the logstore.

Clustering using Brainstorm Optimization

Applied Brainstorm optimization, a swarm intelligence technique, to find optimal cluster clusters in protein dataset. Used Dunn index and DB index for internal cluster validity indices.

Smart Gui for linux terminal

Developed a smart terminal GUI for Debian Linux Kernel 3.14-kali. Created new commands using linux system calls. Stored the user’s history to filter top five frequently used commands and displayed it on the GUI.

Text search using Trie

Developed a cross-platform desktop app in C++ and Qt for searching text in a given paragraph using Trie data structure.

KUcart Ecommerce

A simple E-commerce website.