Projects

TinyML - Machine Learning for Edge-Devices

Documentation

Investigated the toolchains for deploying the machine learning frameworks on edge-devices including the resource-constrained FPGAs. We also designed the FPGA hardware block for overcoming the limited high-speed memory and available data bandwidth while comparing its performance with other embedded devices using a benchmark application. Once the toolchain has been developed and tested, the idea is to implement an application on a suitable hardware by utilizing the tested methodologies. Thus, we are currently evaluating applications in various domains including Genomics for leveraging the developed toolchain.

Sparse Recovery Algorithm on a Low-end FPGA

Source Code Presentation

Implementation of Sparse Recovery Algorithm (Orthogonal Matching Pursuit) on a resource-constrained FPGA. This project has wide applications in areas such as Wide Field magnetic field imaging, Photography, Network tomography, etc.

FPGA Based Emoji Detector

Documentation

Secured 1st position in the GoPynq Competition, organized by Xilinx during Techfest'19. This required interfacing various components including web-cam with FPGA, essential for detection and processing while implementing algorithms such as PCA for Image segmentation and Emoji detection. The final round involved submission of a working protoype, competing with 30 other colleges across India.

Research Project: P-Quest Lab


Working on developing an FPGA based Feedback System and Control Protocols for the Spin-based Quantum Systems, primarily focussing on NV centers, ODMR and NMR experiments for high sensitivity magnetic sensing.

Ultra-fast Current-Voltage (I-V) curve scanner

Presentation User Manual Source Code

Developed and successfully tested an I-V tracer circuit for power semiconductor devices using pulsed I-V measurement technique. The module would be deployed for accelerated testing of solar panels, monitoring the health of by-pass diodes, in manufacturing lines and as testing measurement equipment for various purposes.

Control using Brainwaves

Documentation

Developed a single-channel low-cost circuit for Electroencephalographic(EEG) signal acquisition. This required designing a low-power noiseless filter bank to extract the frequencies of interest (8Hz-40Hz) and analyzing various techniques for feature extraction and classification of acquired signals.

Smart Paper Cutter


A Cost-effective and Environment-friendly cutter which segregates unused portion of the used paper, and thus capable of saving Rs 54k annually by reusing old papers with a peak power consumption of 12W. This was developed as a part of Inter IIT Tech meet 2k18, where we stood as the 2nd runner-up among 23 other IITs.

Smart Vending machine

Documentation

Implemented a cheap, power efficient and user friendly vending machine with safety features including sending sms to the vendor, informing about exhaustion of products or an attempt to damage the machine.

Static Timing Analysis Model


The model computes the execution time of a given set of low-level instruction codes for a class of register-to-register Vector Processors, following in-order execution. This also covers all the dependencies and constraints existing among various instructions.

Pipelined Processor Design


Developed a 6-stage Pipelined processor using VHDL, which was capable of executing 15 instruction sets using an 8-register and 16-bit computing system. This also involved optimizing the architecture performance by including Hazard mitigation blocks and Branch predictors.

Star Tracker-based Attitude Determination System

Abstract

A CubeSat-compatible modular Star Tracker-based Attitude Determination System to be tested on PS4-OP by ISRO. Currently developing an FPGA-based framework for storing and processing Star images onboard.

Advitiy


Advitiy is the 2nd student satellite of IIT Bombay, technically advanced and efficient version of the 1st, Pratham. As a member of the Electrical Subsytem, I was involved in the design of power distribution circuit, interfacing with peripherals and implementation of the control algorithm.

CrakX - Transforming your Preparation

Download now! Website

App to help navigate through the latest campus resources to boost placement and internship preparation. This has acquired 5k+ users since its launch, followed by appreciation from IIT Bombay and users across the country.

Robotic Fish


A Robotic fish that was capable of maneuvering underwater. Designed the complete three dimensional model of the fish, which was later 3D printed and a control algorithm was implemented for maneuvering a 3-link manipulator with servo motors as actuators.

Other Projects


A Line follower bot - An arduino based automated line or path following 3-wheeled bot, based on the Proportional–integral derivative control(PID) algorithm.
XLR8 - An arduino based Bluetooth controlled 4-wheeled bot that overcame an obstacle course and passed the XLR8 competition successfully.

Applying AI/ML for Beam Selection in 5G Networks


Developed a robust neural network model for performing the K-top beam selection task to identify the strongest combined 5G channel in a Vechicle-to-Infrastructure scenario. This would drastically reduce the communication overheads and complexity of the embedded design. Various sensor data including the Lidar point clouds, Camera images and GPS readings were preprocessed to train and test various neural networks. The final model obtained an accuracy of 91% in the top-10 beam selection task.

Title Generation using NLP


Fine-tuned a pre-trained T5 Transformer to generate an appropriate title for the input document and article. The training dataset was formed by utilizing the scientific research papers and news articles. The model was further trained using the PyTorch framework. A Bi-LSTM seq2seq encoder-decoder architecture with attention was also implemented using Keras, and its performance was compared with the T5 transformer.

Product Review Sentiment Analysis


Implemented RNN, LSTM, Bi-LSTM, GRU, and Bi-GRU networks for sentiment analysis of Amazon product reviews using Keras framework. The initial dataset incurred class imbalance problem, and thus this involved preprocessing the customer review dataset using the Synonym Replacement technique for data augmentation. Finally, a comparison of obtained accuracy, precision, recall and F1-score was performed for the pre-trained word embeddings of Word2Vec, GloVe and fastText.

Handwritten Expression Solver


A Solver that recognizes handwritten mathematical expression from an image and subsequently evaluates it. This invloved segmentation of digits and symbols from the given image, and developing a 3-layered CNN for recognizing the segmented components.

Publications