Shalini Ragothaman

Implemented a prototype of a wheelchair that can be controlled with the help of voice commands input to the robot through a developed Android app by the means of Bluetooth modules.
The project was aimed to assist those with minimum to zero movements in their wrists or arms that is essential to use joysticks to control their wheelchair. Used an Atmega328p microcontroller.

A simple automatic light controller where the status of the lights are monitored based on the count of the number of people present in the room. Used an ATmega16 microcontroller.

The project aimed at generating electricity when vehicles run over speed breakers with the help of piezoelectric crystals embedded on the speed breakers.

Performed a detailed analysis of various encoding and regression techniques on Kaggle’s categorical encoding dataset and obtained an F1 score of 0.77 for the challenge.

Studied and Implemented a Visual Question Answering System that produces answers to questions relating to medical images. The med-VQA system was implemented for the RAD-VQA Dataset.

Evaluated two models of Recommender Systems – Online Particle Thompson Sampler and a Deep Item Based Collaborative Filter to examine the “cold-start problem” using Matrix Factorization and Monte Carlo methods. Analyzed the hit rate and cumulative gain of the two models using the Pinterest and the MovieLens 1M datasets and identified the challenges involved in designing a personalized online recommender system.

With the help of Microsoft's AirSim, a multirotor an optimal policy was designed for a multirotor to navigate through the various environments present without any collisions. A reward function was designed and a Deep Q network was implemented that assisted the multirotor to navigate while imposing harsher rewards for every collision and better rewards when the goal position was reached ultimately.

A mobile app was developed that was aimed at assisting bicyclists to slipstream behind fast moving objects or buses. The app helped in identifying objects that were slipstream-able and this was achieved by retraining the YOLO network and converting the model to a light-weight model that can be deployed onto a mobile phone using TensorFlow Lite.

A conference style paper that surveyed various definitions for terms such as scalability, complexity, performance metrics and so on was developed throughout the semester to compare current computer systems' architectures and explore the highly complex architecture that is present in the brain to analyse performance metrics and the parallelism that exists in current computer systems.

An automated system using Deep Learning and Image processing was designed and built that aims at segregating waste without any human interference to prevent degradation of the environment and to relieve of human efforts in the segregation process.
- A deep learning model was created to classify the various images of the waste objects entering the waste sorter system in real time.
- The automated sorting system was developed by building a conveyor belt system that would input the waste… Show more

An automated system using Deep Learning and Image processing was designed and built that aims at segregating waste without any human interference to prevent degradation of the environment and to relieve of human efforts in the segregation process.
- A deep learning model was created to classify the various images of the waste objects entering the waste sorter system in real time.
- The automated sorting system was developed by building a conveyor belt system that would input the waste objects and based on the label with which each object is associated, the waste objects were physically categorized into the various bins, with the entire process being automated and segregated in less than 30s per waste object.
- Performed in-depth studies to analyse the recyclable properties of waste items and which of these objects meet global standards of recycling.
- Achieved an accuracy of 75% when implementing the real-time automated waste sorter system. Show less