Predicting Airbnb customers’ first booking based on their online activity

Implemented blended ensemble of classifiers technique to predict new customers’ first booking. Blended the Logistic Regression, SVM, KNN and XG Boosting models into the ensemble. Increased accuracy by 6%

Spam Detection Using Machine Learning

Detecting spam e-mails. Preprocessed and vectorized an email body content using 'bag-of-words' and 'concentration based feature extraction' techniques, and trained a Neural Network to classify spam mails from legitimate ones. Gained classification accuracu of around 87%

Face Recognition using D-KSVD

Trained a Discriminative K-SVD model, which is a dictionary-based learning model to perform face recognition. Gained a recognition accuracy of 95%

Predict Testing Time of Car

Developed a Two-Level Machine Learning model to predict testing time of a car based on 370 features. Reduced the dimensionality to 125 features and trained the model which predicts with maximum error of 6 secs

Predicting Lung Cancer

Developed a Deep Learning model using 3D VGGNet Convolutional Neural Network to predict a potential Lung Cancer case in future based on the current Lung CT scans. Gained accuracy of 71%

Object Detection and Localization

Trained YOLO algorithm using pretrained VGG16 Convolutional Neural Network to detect objects in an image and Localized them

Nuclei instance segementation

Using Mask-RCNN technique to identify the region of nuclei presence in an image

Sentiment Analysis on IMDB Movie Reviews

Trained a Recurrent Neural Network to detect sentiment of an IMDB Movie Review

Conversational AI Design

Designed a comprehensive Conversational AI system which is a combination of Machine Learning, Sentiment Analysis, Context Analysis and Reinforcement Learning

Navigation by avoiding obstacles using Deep RL

Training a DQN to guide the agent to move around by avoiding obstacles in the path

Optimal Drone Swarm Navigation for Search and Rescue

Developed a planning algorithm for optimal navigation of a swarm of drones to effectively complete search operations in minimal time, identifying risk zones on the fly

Above is the behavior of swarm of drones' navigation across a 64x64 grid area. A Heatmap is randomly generated and is displayed on the right.

At each timestep, drones take an action based on a heuristic function results. Each drone considers the heatmap at that timestep and fellow drone locations as heuristics.

There is balance between exploiting the high risk zones(red zones) and also exploring the unexplored areas for evidences (Humans).

As drones rely on heat map, drones adapt their course in accordance to the dynamic updates in the heat-map.

The blue dots are the places where humans are present but the swarm or base station doesnt know their presence initially. The heatmap is updated once any drone explores and finds a human to make that region more important (red zone).

there is an attraction factor between drones. They tend to get closer when available to keep mesh network intact.