Hi!
I’m Sourish  

My ProJECTS

Skills & Expertise

Data Visualizations

A picture says more than a thousand words..but what exactly do you want to visualize?I have seen a lot of examples of very good bus also very bad visualizations. For example a pie chart with more than 3 or 4 categories is not readable in my opinion. Because people are better at comparing lengths than areas, a bar chart is a better choice in many cases.
I am using rainbow colors only occasionally since there is no “greater than” ordering in colors so it’s hard for a human to compare quantitative data. Until now I have used Matplotlib, Seaborn, Tableau and plotly for my data visualizations with most experience in the first 3.

Machine Learning

Need to make predictions based on historical data? Or want to explore your data for certain patterns?Machine learning is a discipline that can be applied to a lot of domains.
Many companies are working on self driving cars, which are all based on some sort of machine learning.I have experience with supervised, unsupervised and natural Language processing. In one of my recent projects I have for example building a feature matrix for user's historical data from twitter and developed a recommander system based on their proximity.

Web Developement

It's the work that happens behind the scenes to make a website look great, work fast and perform well with a seamless user experience. I have an experience in designing and developing various websites and web based applications using react framework for front end with node.js as as backend server. Interactive data management applications is one of the key areas of my expertise .Quick assist is one of my recent web based application project which helps to connect user with the home service advisers with a seamless video calling facilities.

Portfolio

Project Reports

Texting Behaviour
Feature Engineering | NLP | Modelling

Build a feature matrix for users based on their texting behaviour using NLP and clustering algorithm and developed a recommender system model to identify the proximity of people based on the matrix

Public Health-1
Exploratory Data Analysis | Feature Engineering

Explored and analyse the data to  to identify the key features and evaluate the statistically significant of the features that affect the presence of West Nile Virus in the city of Chicago, Illinois.

Public Health-2
Feature Engineering | Statistical Modelling

Built a multiple regression model with a mean accuracy score of 81.29% to determine how the independent variables affect the presence of West Nile Virus in the city of Chicago, Illinois

Comcast Telecom
Exploratory Data Analysis | Data Visualizations

Analysed the data of Comcast, an American global telecommunication company, to pin down the reason behind their poor customer service which cost the company over $2.3 million as fine by the regulatory authority

Quick Assist
UX - Design | Full Stack Development

Designed and Developed the responsive application using Node js and React which helps to connect user with the home service advisers with a seamless video calling facilities.

NYC 311 calls
Exploratory Data Analysis | Data Visualizations

Perform a data analysis of New York City 311 call to understand the pattern in the data and also visualize the major complaint types.

Hotel Review - 1
Exploratory Data Analysis | Feature Engineering

Perform feature engineering by applying NLP techniques and  analyse data to understand and extract the pattern in the text data.

Hotel Review - 2
Feature Engineering | NLP | Modelling

Developed Ensemble machine learning models to correctly label the sentiment behind hotel reviews.

Notebooks and Links

NIDDK
Feature Engineering | Exploratory Data Analysis |  Modelling
Texting Behaviour of User
Exploratory Data Analysis | Natural Language Processing | Machine Learning
NYC 311 calls
Exploratory Data Analysis | Data Visualizations
Comcast Telecom
Exploratory Data Analysis | Data Visualizations
Hotel Review - 2
Feature Engineering | Natural Language Processing | Modelling
Hotel Review - 1
Data Analysis | Statistical Analysis | Data Visualisation
Bixi Bikes
Exploratory Data Analysis | Data Visualizations
Public Health
Feature Engineering | Exploratory Data Analysis |  Statistical Modelling
Public Health-2
Feature Engineering | Exploratory Data Analysis | Statistical Modelling

What my Client and Colleague
Says About Me?

Sourish is the most incredible developer I have ever worked with. He has a unique ability of being a full stack developer but also a brilliant UI designer, with a phenomenal understanding of UX. The best example of his work is https://www.quickassist.uk/ which is what Sourish was the sole developer for and I, as co-founder can attest to his ability. His work ethic is unlike anyone I have ever seen, as he simply does not stop working until the task at hand is completed. If Sourish does not immediately know how to resolve an issue, rest assured that he will find out.

The other aspect of Sourish that should be mentioned is his ability to think about what he is developing from a practical business sense and what works and what doesn't and what helps to push a business forward. I worked effectively 7 days a week with Sourish for almost a year with Quick Assist and every day working with Sourish he demonstrated an incredible level of integrity and honesty, which was very important throughout especially when things didn't always go to plan.

This was further made easier by his very easy going and personable nature.Anybody working with Sourish would be extremely lucky to have such an incredible driven, inquisitive and talented individual.
Miles Szabo
Co-Founder, Quick Assist
It was a privilege to work with Sourish at BrainStation - both in group work and in our Agile Pod. He is very thoughtful and dives deep into every subject so he can understand it clearly. Whenever I had a question about why something was the way it was or what his thought process was, he took the time to thoroughly explain it in detail and answer any follow-up questions I had. Seeing his work on his individual project was inspiring and thought-provoking.

Not only has Sourish been an excellent classmate, but he was also considerate of my personal growth and helped me improve my own individual project out of the goodness of his heart. With Sourish's wealth of knowledge and social skills, I know any team would be lucky to have him.
Autumn Salasberry
BrainStation Colleague

About Me

Being a professional designer-developer and having more than 7 years of experience in academia, I had the opportunity to explore the close relationship between changing human behaviour and trends in the data. Additionally, my interest to seek for patterns in understanding a problem and weaving the same in the form of a solution has been instrumental in extending my knowledge and driving my career in DataScience.

My recent training in Data Science at Brainstation has given me a breadth of technical skills in Python, SQL, ML frameworks and big data analytics. Further, my prior experience of freelance practice has helped me in building a strong foundation to generate industry-oriented insights and develop data-driven solutions for stakeholders.

I am excited to leverage my recent training and my past experiences to contribute and explore the future challenges in the field of Data science. Please feel free to reach out if my profile resonates with you, or in case you have any questions about my projects, data, or anything in general.

RESUME
Copyright © 2022 - All Rights Reserved
x
In the two part study, the first part aims to explore, understand and analyse the basic pattern of the spread of the mosquito using the tracking data and in the second part ,the study drive deep into analysing the relationship between the different feature variables and building statical models to explore how the independent variables affect the number of mosquitoes caught and the presence of West Nile Virus .
Context
‍West Nile Virus (WNV) is a viral illness largely spread by mosquitoes. The disease is transmitted to a person when an infected mosquito bites them. The city of Chicago, Illinois has been keeping track of mosquito population and WNV prevalence using a series of traps that they place around the city.

Analysis
The study inferred that the maximum possibility for the disease to spread by the mosquitoes are within the 27th-37th weeks of the year as the count is above 2000 for both the species i.e. CULEX RESTUANS and CULEX PIPIENS who are predominant carriers of the WNV.
‍
From the statistical analysis using linear regression model and correlation matrix , for exploring the relationship between features, it was observed that there are strong relationship between mosquito species and the occurrence of West Nile Virus. Furthermore the species of the mosquito , the time of the year (month) and the trap type are the three most important factor for the presence of the WNV in the mosquitoes trapped.
‍
However the analysis ended with the conclusion that the number of mosquitoes trapped don't have a very significant impact over the WNV  as the odd-ratio is close to zero at the optimal threshold level of 0.6 for the logistic regression model.
‍
GitHub
x
West Nile Virus (WNV) -II
This part of the study drive deep into analysing the relationship between the different feature variables and building statical models to explore how the independent variables affect the number of mosquitoes caught and the presence of West Nile Virus .
Analysis
From the initial statistical analysis using linear regression model and correlation matrix , for exploring the relationship between features, it was observed that there are strong relationship between mosquito species and the occurrence of West Nile Virus. Furthermore the species of the mosquito , the time of the year (month) and the trap type are the three most important factor for the presence of the WNV in the mosquitoes trapped.
However the analysis ended with the conclusion that the number of mosquitoes trapped don't have a very significant impact over the WNV  as the odd-ratio is close to zero at the optimal threshold level of 0.6 for the logistic regression model.
‍
GitHub
Comcast Telecommunication  
Context
‍Comcast is an American global telecommunication company. The firm has been providing terrible customer service. They continue to fall short despite repeated promises to improve. Only last month (October 2016) the authority fined them a $2.3 million, after receiving over 1000 consumer complaints.

Analysis
From the initial data analysis to it was explored that the state of Georgia has the number of complaints and the unresolved complaint is also maximum with close to 15.48 % of the overall open cases. The study ended with the conclusion that company need to spend more resources in uplifting the quality of the customer care service as 77.21% cases are solved which were received via Customer Care Call and 76.29% solved which were received via telephone
‍
x
NYC-311 Calls
The projects aims in data wrangling techniques to understand the pattern in the data and also visualize the major complaint types
Context
‍NYC 311's mission is to provide the public with quick and easy access to all New York City government services and information while offering the best customer service. Each day, NYC311 receives thousands of requests related to several hundred types of non-emergency services, including noise complaints, plumbing issues, and illegally parked cars.

Analysis
After analysing the data it was observed that New York, Brooklyn and Bronx city are three major cities form where the most number of complains calls are made and noise from the street and sidewalks ,the illegal Parking and Blocked Driveway are the major cause for it. However at the end it was observed that there where no statistical significance in the relation between the complain type and the location [city]
‍
GitHub
x
Hotel Review Analysis
The study is about the Exploratory Data Analysis (EDA), and then move into data augmentation, modelling, and iteration over model improvements.
Analysis
From the EDA , it was inferred that the overall distribution of the Reviewer Score spread between 2 -10 points. Close to 80% of the reviewer rated the hotels in the range of 8-10 points which make the over all distribution skewed towards the higher score and among the top score, 36.40% of the reviewer rated the hotel with full 10 points.This makes the over dataset highly imbalanced while considering the Reviewer Score as the target variable. Therefore converting the Reviewer_Score into binary columns ,the overall dataset turned to be more balanced.
‍
‍
GitHub
x
Hotel Review - II
To develop several machine learning models to correctly label the sentiment behind hotel reviews.
Analysis
From the EDA , it was inferred that the overall distribution of the Reviewer Score spread between 2 -10 points. Close to 80% of the reviewer rated the hotels in the range of 8-10 points which make the over all distribution skewed towards the higher score and among the top score, 36.40% of the reviewer rated the hotel with full 10 points.This makes the over dataset highly imbalanced while considering the Reviewer Score as the target variable. Therefore converting the Reviewer_Score into binary columns ,the overall dataset turned to be more balanced.
‍
GitHub
Context
Social media platforms  based on their account suggestions policy, clearly give more priority to the way users like various contents or share them with others, or on the common geographic location of the users etc. This approach undermines the fact that texting (which are in public domain) in itself is a significant touch point for the platform and as a result, these platforms sometimes fail to connect like-minded people

Business Value
This recommender system model, will not only help to create the possibility of the like-minded people to get connected, but also provide them the opportunity to build their own community for exchanging thoughts and ideas on the topics of their interest. From the business perspective, the model will help companies to closely identify their potential customers for cognitive marketing so that ad-effectiveness is improved and resource allocation is optimized.

About the Data
The study is conducted using the historical text data from twitter. The actual data used for the project is taken from the study titled ‘Sentiment140 dataset with 1.6 million tweets' published on the website ‘www.kaggel.com’.
‍
Methodology
‍The study was conducted in three stages. In the first stage of “Feature Identification” process feature engineering was done, with the goal to evaluate
1. the top level behaviour of texting like sentiment analysis, frequency of texting and usage of graphical support,
2. topic level inclination while texting using Latent Dirichlet Allocation algorithm
3. lexical semantics of texting using word embedding techniques like word2vec and PCA (to handle high dimensionality).
In the second stage, the feature matrices created in the previous stage were used to develop the recommender system with the support of Content based modelling.
In the final stage, an algorithm was developed using clustering algorithm like Kmean and DBSCAN and graph network which can validate the results of the recommender system model.

Conclusion
Based on all the exploration, analysis and modeling, it is observed that texting behaviour is a very complex aspect of user’s interaction on any social media platform. As per the results of all the three models on the given dataset, it is clear that each feature matrix represents a different set of user texting behaviour, which are independent of each other. Furthermore, specifically the topic level and text level model also support the initial hypothesis that texting behaviour does not reflect the users profession or the background completely, as the user recommended are not very conventional. This overall opens the avenues for the model to recommend new people to connect with an entirely different approach. Lastly, the feature matrices along with the model outcomes can act as a guiding tool for the user to assess their content and style of texting.
‍
x
Bixi Bikes
The goal was to gain a high level understanding of how people use Bixi bikes, what factors influence the volume of usage, popular stations, and overall business growth.
Context
BIXI Montréal is a non-profit organization created in 2014 by the city of Montreal to manage its bike-sharing system. The BIXI network has more than 9,600 bikes and almost 800 stations spread out across the areas of Montreal, Laval, Longueuil, Westmount, Ville Mont-Royal and Montréal-Est.

Analysis
From the initial exploration of the data, it was observed that July and August are observed as the peak time where not only the total trip count is maximum but also the average number of trips per day.The other significant factor, was that the members’ contributions were potentially higher as compared to the non-members across the season in terms of the total duration(in seconds) of usage. In the final stage of the study it was observed that the primary reason for  stations having maximum number of roundtrips is because of its remote location.
‍
GitHub
x
National Institute of Diabetes and Digestive and Kidney Diseases
The objective was to predict whether or not a patient has diabetes, based on certain diagnostic measurements included in the dataset
Context
The mission of NIDDK is to conduct and support medical research and research training and to disseminate science-based information on diabetes and other endocrine and metabolic diseases; digestive diseases, nutritional disorders, and obesity; and kidney, urologic, and haematological diseases, to improve people’s health and quality of life.

Analysis
In the initial level of the analysis it was observed that patient with high Glucose level and insulin level are more prone to diabetics. Furthermore patient with the skin-thickness above 25mm and body mass index above the 40 unit are most prone to be diabetic. The Insulin level is the key reason for diabetes as the data clearly indicates that above the level of 80-100 units, most patients are diabetic. The similar observation can be seen in the age data where at the older age women suffer from diabetes.
For the final prediction, the XB boost model was considered  with the optimal threshold level at 0.195 with a recall score(which is considered as the primary metric for evaluation) of 87%
‍‍‍
GitHub