Projects

Working with Rajamani Ayyer, this was our Major Project for the B.Tech. program in Electronic & Communication Engineering. Due to the Covid19 pandemic, working remotely on the project was tough but fortunately we were able to complete the main functionality & present it virtually.

The main aim of the project was to develop an IoT-based 'Smart Hospital' environment. We wanted to integrate hardware and software in medical applications which would benefit doctors & nurses. Incorporating different fields of technology, we wanted to build something more valuable than its parts.

At the center of our project was a web server we created using Django. We also developed a Patient Health Monitoring system which shared patient vitals to the server through MQTT. The vital measurements included Body Temperature, SpO2, Heart Rate & ECG. Our main idea was to embed the different sensors into the patient's hospital gown to help nurses continuously, conveniently & remotely monitor their health. We also included simple thresholds to alert nurses in case of abnormally high or low vitals.

We developed a website from scratch for patients to check on their health and for nurses & doctors to monitor several patients' health in one glance. We also added a platform for Electronic Health Records for doctors to make digital prescriptions for a patient. Further, we developed an interactive data analytics dashboard using Plotly with statistics like room occupancy, staff gender ratio & patient demographics for the hospital administration to gain insights and help with the decision-making process.

We also developed an ECG display which can read, process, visualize and analyze a patient's ECG waveform using Digital Signal Processing. We also trained an Multi-layered LSTM to identify abnormalities in the heartbeat pattern. Additionally, we also developed a simple 'Smart Sanitizer' which automatically alerts any individual entering into a patient's room to sanitize their hands. This is a very recommended practice in hopsitals, especially in ICUs & OTs to prevent cross-contamination.

For more information, you can watch the video of a working demonstration of our project below -

This was my submission for the Cube Highways Hackathon 2021. After an idea screening round, the top 20 finalists had to complete a working prototype in less than a week. The main idea of my proposal was to create an online platform where travelers would be able to pre-order food before actually entering a restaurant. This would save the travelers time on their journey and enable the restaurant to improve their service rate. To make the experience more convenient & memorable, I also applied a few Machine Learning models.

The frontend was developed from scratch using HTML, CSS, a bit of JavaScript and with the help of BootStrap. The web server was developed in Django. Travelers can view the Virtual Menu to look at the price, description, images & even others' feedback about different food items. While ordering, they can apply coupons & after ordering, they stand a chance to win coupons.

I also added a Quick Feedback option for customers to give extremely convenient feedback about different food items - simply take a picture of the food & add a star rating. The customer does not have to mention the food item for which the rating is being given, since the food item is automatically identified by my ML model using the picture.

For more information, you can watch the demonstration video of my project -

This was my submission for the Worldwide AI Hackathon 2021. The main idea of Signalytics is that it is a simulation & analytics tool for testing different kinds of dynamic traffic signal scheduling algorithms for various traffic conditions.

Simulation setting can be made for complex traffic lighting system at intersection. The simulation has realistic features in car movement at intersections, acceleration/deceleration, etc. It also has the capability to spawn emergency vehicles and pedestrian call button triggers at given probability rate. The simulation tool provides total number of cars on a lane, not total number of cars on a lane going in a particular direction, to the scheduler; thus mimicking actual traffic signal cameras.

Simulation data is recorded over a period of 24 hrs simulation time after which different algorithms can be tested based on unique metrics like avg waiting time, avg emergency vehicle waiting time, traffic lane length, pedestrian waiting time, etc.

For more information, you can watch the demonstration video of my project -

This was my submission for the Tensorflow Microcontroller Challenge. The main objective of the challenge was to create something fun which highlighted the application of Tensorflow Lite for Microcontrollers.

My main idea was to recognize handwritten digits by tracking the pen's motion. Since we can expect the pen to move in the same shape as whatever digit is being written, we can simply track the pen's motion to predict digits without the need for cameras or other visual sensors. Through a simple Tkinter User Interface in Python, this project essentially digitized pen-and-paper writing. I also attached piezo-electric cells underneath the writing surface to ensure digit recognition happens only when the pen is actually writing on the board.

I also added an extremely simple Dim Lighting Assistance to help write in darkness while also adding another TinyML model for gesture recognition wherein one can customize the cursor position, font color & font size through gestures.

For more information, you can watch this short & fun demonstration video -

This was my submission for the India PropTech Hackathon 2021. After an idea screening round, the top 20 finalists had to complete a working prototype in around a week. The main idea of my proposal was to create a ChatBot which is aimed at providing reassurance to homebuyers by answering their personal queries and updating them about the construction progress. This helps in making the construction details transparent and improving post-sale customer services. I have used different Machine Learning concepts like Natural-Language-to-SQL-Query conversion & Time Series Forecasting for this project.

I developed a simple graphical user interface for the chatbot using Tkinter in Python. The construction progress is visualized to the homebuyers through a dashboard detailing the completion percentage & forecasted date of completion. Forecasts have been made using the latest progress of the given construction and also data from previous constructions.

To make the chatbot capable of answering personal questions, I have incorporated NL-to-SQL concepts. I have used the pre-trained SQLNet model along with developing some additional code & tweaks. The SQLNet model is capable of converting the user's question into an SQL query which can then be used to fetch the answer from the database.

For more information, you can watch the demonstration video of my project -

I have made it a personal goal to regularly work on different Kaggle datasets about different topics & containing different kinds of data. It help me in not only finding new ways to tackle a problem, but also in revising basic data analysis concepts and my Python & R programming skills. I have hid the coding part of most analysis (it is available on Kaggle, if interested) to let viewers quickly glance over the main insights without wasting much of their time. I hope people have as much fun discovering interesting facts as I had uncovering them!

Working with Sahith S R, this was our second Mini-Project for the B.Tech. program in Electronics & Communication Engineering at NITK Surathkal. The objective of this project was to combine several standard Deep Learning architectures to achieve best performance in Teeth Segmentation.

We implemented the teeth segmentation on Dental X-rays using several models like Fully Convolutional Network (FCN), Standard U-Net, Attention U-Net, High-resolution Encoder-Decoder Network (HMEDN), ResNet, DicNet, Atrous Spatial Pyramid Pooling and several combinations of these models. Eventually, we observed that models like HMEDN & model combinations like ResNet+DicNet+ASPP performed much better than standard image processing techniques, giving almost 12% increase in Dice score, 3.5% increase in Accuracy and comparable IoU scores.

This was a course project that I worked on with Kumar Alabhya, Sachin Doddamani & Vishal Ray. Our objective was identifying the speaker's emotion from their audio speech signal using Deep Learning models.

Our approach involved extracting temporal, frequency & statistical features from the audio signal. Features like frame-wise Energy, frame-wise Zero-Crossing-Rate, Mel-spectrogram, MFCCs, LPCs, HNR, etc. were extracted and passed through a Multi-path Neural Network of Dense layers & Conv-LSTM layer with Local Attention. The model was tested using the RAVDESS & IEMOCAP datasets.

I had a lot of fun building this website as one of my pet project. My main priorities while designing this website were mobile-responsive content which is quick to glance for the viewers. I enjoyed customizing the design, colors & content of the website, which was a pretty satisfying experience. I wanted this website to be a virtual portfolio of my work and a digital resume for viewers. And I feel quite proud to be able to call this microscopic website in the vastness of the internet my own space! 😍

Working with Arpit Jadiya & Arpitha R., this project was aimed at adding new features to IRIS, NITK's 'Integrated Resource & Information Sharing' website through its research division, IRIS Labs. The main idea was to enable Natural Language to SQL query conversion in the website's chatbot so that the bot can efficiently search the IRIS database for answers. The main steps involved firstly processing the question text, then identifying the general topic of question, then guessing the table in the database which might contain the answer & finally applying a pre-trained NL-to-SQL model like SQLNet to generate an SQL query.

One of the club projects at IEEE NITK, I worked with Tulasi Chandana, Akshara P., Shruthan R. & K. Krishna Swaroop for this project. The project involved detecting sarcastic comments in Reddit, using not only the content-based information (the textual content), but also context-based information (the topic of discussion and the user's comment history). We also used BERT as an alternative to creating sentence embeddings.

This was my first Mini-Project as part of NITK's curriculum. Working with Vishal Ray, we coded in basic MATLAB and applied Digital Signal Processing concepts to process & visualize a Synthetic Aperture Radar's output.