Varun Desai | EE, IIT Delhi

Internships

APT Portfolio Pvt. Ltd., Bangalore.

RTL Design and Verification, FPGA-based Simulation Acceleration

Ideated, designed, tested architecture for emulating RTL model on an FPGA to speed up verification process more than 100x
Developed and tested digital circuit using Verilog to (de)serialize data to and from compressed format for efficient transmission
Used Riviera-PRO and Cocotb to debug and test design, Xilinx Vivado for synthesis and ensuring timing requirements

Greenleap Robotics, Delhi

Automated Fault Diagnosis of Solar Panels

Implemented Machine Learning based Computer Vision techniques along with statistical methods to automate fault diagnosis
Trained a ConvNet with state of the art architecture(Mask-RCNN) to localize solar panels in aerial drone-shot thermal images
Used Image Processing and Unsupervised Learning(DBSCAN algorithm) to detect temperature hotspots as statistical outlierss
Achieved over 90% precision and recall on test dataset; developed simple frontend using python libraries such as tkinter

Projects

Analong ConvNet

Prof.Debanjan Bhowmik

Matrix multiplication occurs in O(1) time in a crossbar array but with volatile transistors various non-idealities hinder scalability
Introduced weight decay of the synaptic device while training by mathematically modelling the decay curve; used tensorflow
Experimented with different quantization precision, rate of decay, and distributing the weight between volatile and non-volatile devices

Navigo

Winner, i4 2020 at IIT Delhi

Capacitated Vehicle Routing formulated as an Integer Programming problem; Researched existing solutions and improved upon them
Introduced Developed attractive, aesthetic front-end for various stakeholders; highlights path taken for every agent; used JS, Flask
Awarded best solution after evaluation by professors & experts in the field, out of 70+ submissions for 28 proposed problems

Code can be found on github.

RL, MDPs, and Kalman Filters

Prof. Rohan Paul, Special Topics in AI

Implemented value iteration for solving an MDP, reinforcement learning algorithms(SARSA,Q-learning); Viterbi Algorithm for HMMs
Devised algorithm for estimating position of multiple agents with no data association as an extension to Kalman Filtering

Code can be found on github.

Machine Learning and Intelligence

Prof. Prathosh AP

EM for GMM implemented, Used PCA for dimensionality reduction on Medical MNIST dataset, reduced convergence time
Neural Net(L1 and L2 regularization), LeNet, AlexNet, VGG-16 trained with Adam Optimizer using PyTorch on HPC, IITD

Code can be found on github.

Audio Source Separation

Prof.Lalan Kumar

Working on algorithm which uses machine learning concepts to perform Independent Component Analysis and Principal Component Analysis to separate a mixed audio signal into two different audio signals corresponding to different sources.
Code implementation will be done on MATLAB/Python

Data Structures Course Projects

Prof.Subodh Kumar

Implemented Double Hashing, and separate chaining using binary search trees from scratch on Java
Implemented a Job Scheduler, which stores jobs in a MAXHEAP and executes them on basis of it’s projects priority
Implemented Graph Data Structure in Java to store the relative orientation of various triangles in the Cartesian Coordinate System

Histogram Based Image Smoothing

Prof.Abhishek Dixit

Built a program which processes an image’s histogram and tries to figure out which enhancement technique works best: Histogram Equalization or Contrast Enhancement and applies them with the right weight. Code was implemented on Octave/ MATLAB

Arduino Music System

Independent Project

Built a system using an 8*8 LED matrix, shift registers, a simple microphone and an Arduino Uno, which generates an approximate real time amplitude-time digram on the matrix for input audio signal
Can be used while listening to music, gives a nice pattern on LED Matrix
Link to Video