Sanidhya Mangal

mangalsanidhya19@gmail.com // CV // Scholar // github // twitter // Linkedin

Sanidhya Mangal is a highly skilled Data Scientist with a strong background in computer science and over two years of professional experience in data analysis, machine learning, and AI deployment. Currently working at HealthEdge, Sanidhya has made significant contributions, including deploying a GenAI chatbot that enhanced customer satisfaction and reduced ticket resolution time by 38%. Sanidhya's technical expertise spans across Python, SQL, Docker, Kubernetes, and various AI tools, enabling him to optimize ETL pipelines, automate processes, and develop models that save millions in operational costs.

With a Master’s degree in Computer Science from Vanderbilt University and a B.Tech from Medi-Caps University , Sanidhya has a solid foundation in data science and machine learning. His experience includes developing advanced ML models, leading projects that improve efficiency and accuracy, and contributing to research with publications in genome filtering and cancer diagnosis. Sanidhya’s portfolio showcases a blend of technical proficiency, innovative problem-solving, and impactful project outcomes, making him a valuable asset in the field of data science and AI.

February 2023: Commenced my early career with HealthEdge as Associate Data Analyst
December 2022: Completed my MS in CS degree from Vanderbilt University with 3.94 GPA
August 2022: Automated filtering of genome-wide large deletions through an ensemble deep learning framework got accepted in Methods journal.
August 2022: Wrapped my 12 Weeks of internship @ Asurion.
November 2021: An ensemble deep learning framework to refine large deletions in linked-reads paper got accepted in BIBM'21 conference.
October 2021: Won Silver Place @ VandyHacks VIII.
August 2021: Joined Vanderbilt University under MS in CS program.
January 2021: Watch an exclusive interview describing what Machine Learning Engineer's do at Engineerbabu.

publications

Automated filtering of genome-wide large deletions through an ensemble deep learning framework Yunfei Hu, Sanidhya Mangal, Lu Zhang, Xin Zhou
Methods
bibtex

@article{HU202277,
	title = {Automated filtering of genome-wide large deletions through an ensemble deep learning framework},
	journal = {Methods},
	volume = {206},
	pages = {77-86},
	year = {2022},
	issn = {1046-2023},
	doi = {https://doi.org/10.1016/j.ymeth.2022.08.001},
	url = {https://www.sciencedirect.com/science/article/pii/S1046202322001712},
	author = {Yunfei Hu and Sanidhya Mangal and Lu Zhang and Xin Zhou},
	}

An ensemble deep learning framework to refine large deletions in linked-reads Yunfei Hu, Sanidhya V Mangal, Lu Zhang, Xin Zhou
bioRxiv
bibtex

@article{hu2021ensemble,
title={An ensemble deep learning framework to refine large deletions in linked-reads},
author={Hu, Yunfei and Mangal, Sanidhya V and Zhang, Lu and Zhou, Xin},
journal={bioRxiv},
year={2021},
publisher={Cold Spring Harbor Laboratory}
							  }

Convolution Neural Networks for diagnosing colon and lung cancer histopathological images
Sanidhya Mangal, Aanchal Chaurasia and Ayush Khajanchi.
arXiv

code // bibtex

@misc{mangal2020convolution,
title={Convolution Neural Networks for diagnosing colon and lung cancer histopathological images},
author={Sanidhya Mangal and Aanchal Chaurasia and Ayush Khajanchi},
year={2020},
eprint={2009.03878},
archivePrefix={arXiv},
primaryClass={eess.IV}
}

Lstm vs. gru vs. bidirectional rnn for script generation
Sanidhya Mangal, Rahul Modak and Poorva Joshi.
arXiv

code // bibtex

@misc{mangal2019lstm,
title={LSTM vs. GRU vs. Bidirectional RNN for script generation},
author={Sanidhya Mangal and Poorva Joshi and Rahul Modak},
year={2019},
eprint={1908.04332},
archivePrefix={arXiv},
primaryClass={cs.CL}
}

LSTM Based Music Generation System
Sanidhya Mangal, Rahul Modak and Poorva Joshi.
IARJSET

code // bibtex

@article{Mangal_2019,
title={LSTM Based Music Generation System},
volume={6},
ISSN={2393-8021},
url={http://dx.doi.org/10.17148/IARJSET.2019.6508},
DOI={10.17148/iarjset.2019.6508},
number={5},
journal={IARJSET},
publisher={Tejass Publisheers},
author={Mangal, Sanidhya and Modak, Rahul and Joshi, Poorva},
year={2019},
month={May},
pages={47–54}
}

Projects

SSLGAN

This project is an implementation of inducing Self-Supervised learning based approaches with GANs to improve the training of GANs and prevent forgetfull discriminator

Unsupervised rotation and contrastive learning based methods were used as self-supervised learning methods, on anime face images.

We obtained better FID score for self-supervised contrastive methods. A complete report could be found at this link.
AdvAttacks

This project is an implementation of targeted and untargeted adversarial attacks on imagenet model using FGSM and PGD methods

This repo also contains a pre-training based represenatational learning based approaches to learn how feature learner learns ebmeddings.
Patient Notes Scorer

Machine learning based approach to extract the features from the patient notes recorded by the candidates taking medical exam

Three different models were tested, Logistic Regression, Decision Tree and Naive Bayes. TF-IDF algorithm was used to preprocess text into feature vectors.

Note: BERT based models were also used to perform char level annotations where Ro-BERTa based implementation provided the stellar results
Word2vec

Implementation of Word2vec using negative sampling method on two different data extracted from Reddit.

It contains both the implementation, abstract one using gensim and low level using Pytorch.
PAC

Implementation of Suprisingly Simple Semi-Supervised Learning on office dataset.

Various representation learning based methods such as unsupervised rotation and supervised contrastive learning on the different variations of semi supervised and unsupervised learning.

ResNet-50 was used as a backbone feature extractor, despite of implementing the exact method I failed to reproduce the results mentioned in the paper. A complete report could be found at this link.
DeepMCSolver

Project implementing DDPG and reinforce algorithm to solve the mountain car enviornment.

The entire implementation is environment agonistic i.e., could be used with any continous environment. Therefore it DDPG was tested on LunarLander problem too.
Deep SV Filters

Project to detect deletion structural variants (SVs) in chromosome filtered by Aquila.

To achieve state of the art results various deep learning models were used and experimented. The entire training procedure is backed by transfer learning.

The model generates a BED file as an output of prediction workflow which could be later compared with geneome reference using truvari
GANs

A light weight framework for tooling Generative Adversarial Networks(GANs).

It is built on top of TensorFlow supporting Keras and Eager Execution. Currently it supports data pipelines, convolution models, multiple flavours of GANs

Note: This is a pre-release in an alpha mode. Constant efforts are made to enchance it and add more architectures from GANs
Animator

A NN for generating faces of Anime Chars in RGB channel with channel height and width of 64.

This generative model follows unsupervised generative strategy backed by Least Squares Generative Adversarial Network (LSGAN) model
LSGAN for Fashion Mnist

A trained Least Squares Generative Adversarial Network (LSGAN) model to generate fashion mnist images.
Cifar Wasserstien GAN

A partially trained deep convolution generative adversarial networks(DCGAN) trained using Wasserstien strategy.

All the implementation was done using keras and tensorflow, it took around 2500 iterations to show some comparabale results.

Note: Due to limitations in compute resources model was only trained to above mentioned epochs. To get better results training can be resumed from current checkpoint.
Wasserstien GANs

Powered by deep convolution generative adversarial networks(DCGAN) and backed by Wasserstien based learning strategy to generate fashion mnist images.

All the implementation was done using keras and tensorflow, it provided comparabale results in around 100 EPOCHS
Denoising Auto Encoder

Powered by Auto Encoders(AE) to denoise fashion mnist image dataset.

All the implementation was done using keras and tensorflow, it provided a stellar results in just 10 EPOCHS

Top two rows display noisy images whereas the last ones redenders reconstructed images corresponding to those noisy images
Fashion Image Generation using DCGAN

Powered by deep convolution generative adversarial networks(DCGAN) to generate fashion mnist images.

All the implementation was done using keras and tensorflow, it provided a stellar results in just 50 EPOCHS
Computer aided diagnosis for Pneumonia CT scan and Malaria blood cell slides

Powered by convolution neural network(CNN) to classify lung CT scans into viral or bacterial pneumonia or normal and detecting malaria pathogens from blood sides.

All the implementation was done using keras achieving an accuracy of 92% and 94% for pneumonia and malaria respectively.

A flask based webapp was developed for serving the models
Covid Screening using lung CT scan images

Powered by convolution neural network(CNN) and backed by transfer learning, MobileNetV2 architecture pretrained on imagenet was used to screen CovSars-2 virus.

Two seperate models were trained one to tune only after 100 layers and other to tune complete architecture with AUC of 0.61 and 0.53 respectively.
Fashion Image generation using VAE

A shallow variational autoencoder for generating fashion images based on fashion mnist which was successful in learning latent information from images
Language Translator: English to Hindi

A shallow copy of GNMT for developing a real time machine translator for translating English language to Hindi. We achieved a BLEU score of 14.2 after successful training of a model.
Image Classification for Paleography Analysis

Inspired by the story of In Codice Ratio an attempt was made to create an image classifier for predicting the medevial texts from image using a shallow CNN.
Generating Long articles like Wikipedia

Developed a LSTM backed model for generating long sequence of text data based on wikidata of text. This project laid a foundation of my research inspiration for comparing various sequence-to-sequence models for generating tv scripts.
Face Detection

A project aimed to detect human face from an image using support vector machines and HOG features. Overall accuracy of 78% was achieved after training a complete model.
Chatbot for Single QnA

A chat bot was developed using bAbI dataset for answering a single contextual question.
Neural Style Transfer for generating artistic images

We used VGG-16 architecture pretrained on imagenet dataset as a base model, and after tuning the model with the custom styles using gram matrix for filtering and calculating style of an image.
Text Classification(Sentiment)

A two layered bidirectional GRU backed model intended to classify polarity of text on IMDB movie reviews.