KeshavAnandCode/resume
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Finished CS section

Browse Source
This commit is contained in:
Keshav Anand
2025-12-08 22:19:11 -06:00
parent 856d3e5046
commit e94b461461
2 changed files with 19 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
rsi-application/main.tex
View File

@@ -220,21 +220,31 @@
\vspace{0.2 cm}
\newcommand{\mytext}{Firstly, I find Computer Science deeply intriguing as it is the perfect crossover between my two passions of
mathematics and problem-solving. The ability to optimize a computer chip to solve a pratical problem has
always fascinated me, and the rising popularity of Machine Learning and Artificial Intelligence further
piqued my interest during the COVID-19 pandemic. My research project from 2024 - 2025, GaitGuardian, started my journey
with signal processing, as I had worked on a project to predict Freezing of Gait (FoG) episodes in Parkinson's
Disease patients using a belt-mounted IMU sensor and machine learning algorithms. In addition to building
an end-to-end hardware embedding with a custom PCB, I also developed a strong understanding of signal processing
techniques. My novel pipeline involved using fourier transforms, z-score normalization, and wavelet denoising
\newcommand{\mytext}{During the antisocial years of the COVID-19 pandemic, my endless YouTube bingeing
led me into
accidentally discovering how computers work at a low level. The moment I learned the fascinating
marvel of engineering that is a computer, I immediately fell in love with Computer Science.
Over the years, I have also developed a specialized interest in signal processing and its applications.
My curiousity in this topic stems from my first Math Club meeting in 9th grade, where a senior officer
had explained the fascinating science of how radio signals are transmitted and received using Fourier transforms
(as an application of calulus).
These concepts overlapped with my learning of fascinating Calculus concepts, and I was amazed at how signals can
be analyzed and processed. Today, the major question that excites me within the field of Signal Processing is how to effectively
adapt digital signal processing and machine learning for real-time resource-constrained embedded systems.
As a hardcore robotics enthusiast, I have always been not just interested in the theoretical software,
but also the practical hardware embedding of these algorithms. This really sparked my interest in the field
of signal processing for embedded systems, prompting me to start a 2 year research project that would become the
focus of my life.
My ISEF-winning research project from 2024 - 2025, GaitGuardian, was my first major experience with signal processing,
as I had worked on a project to predict Freezing of Gait (FoG) episodes in Parkinson's
Disease patients using a belt-mounted IMU sensor and machine learning algorithms. My novel pipeline involved using fourier transforms, z-score normalization, and wavelet denoising
to filter out noise from the raw IMU data. Unlike existing approachs that used time-domain features, I fed the
cleaned time-series data into a 1D CNN, acting as an automatic feature extractor (with no flattening).
This was passed into a hybrid biLISTM with temporal and spatial attention mechanisms,
allowing for segmented windows to be read both forwards and backwards, and a final dense layer would
output the boolean state of whether a FoG episode was occuring in real time. The learning I gained from this research led
me to pursue other related signal processing tasks to boost the final product for Parkinson's patients.
Researching other Parkinsonian symptoms led me to explore tremor detection (uncontroleld shaking of the hands),
Researching other Parkinsonian symptoms led me to explore tremor detection (uncontrolled shaking of the hands),
and I implemented a real-time tremor detection model that involved a bandpass butterworth filter to isolate tremor frequencies
between 4-6 Hz, followed by an FFT to extract frequency-domain features. These features were then fed into a lightweight
1D CNN, resulting in state-of-the-art 99\% accuracy while limiting false positives. I also looked into signal processing