Website Development for an NGO

2020

My team and I worked with Cropvet Agro Pvt Ltd. from to create and deploy a website for them. This organization is an NGO Rural Development Organization for empowering the rural population and enhancing their income. It promotes entrepreneurship for the youth in rural areas of India, helping them socially, economically, and financially. The sample deployment is unfortunately not available anymore.

Tech Stack used:

• Flask
• SQLAlchemy
• HTML5
• CSS
• Bootstrap
• Javascript
• MySQL
• XAMPP

Github link: https://github.com/notgitika/website-for-mini-project

Letter of Appreciation by the CEO:

Snapshots from the website

Website Development for Sahaara Foundation - An NGO for Rehabilitiation Services

2020

After my team and I created a website for an NGO, we were approached by another NGO to revamp their existing website. This organization is a Rehabilitiation center for addicts and mental health patients.

Tech Stack used:

• HTML5
• CSS
• Bootstrap
• Javascript

Github link: https://github.com/manshri99/Sahara-Foundation

Snapshots from the website

Genre Identifier from Song Audio File

2020

My team and I created a website using Flask that employed a machine learning model in the backend which analyzes audio files uploaded by the user to predict what genre they are based on their audio spectrum slices.

Introduction:

• With the current state of music and influence of the internet, music from all parts of the globe is consumed by people which makes music very diverse.
• This makes it harder to categorize an artist or song under one specific genre as most new music is usually a blend of multiple influences from diverse sources.
• Music genre classification is, as described ahead, an ambiguous and subjective task. It is also an area of research that is being contested, either for low classification accuracy or because some say that one is not able to classify genres that do not even have clear definitions
• So we have attempted at making a genre identifier with an accuracy of approximately 82%.

Project Workflow:

• In this web application, the homepage has a choose audio file button which lets us choose files to insert a song for the model to identify its genre.
• The inserted audio file is separated into small slices and each slice is assigned a genre after analyzing them from a CNN (Convolutional Neural Network) model. The final genre is decided from a majority voting system.
• This application supports various audio formats which makes it more accessible for various users. The more popular examples of audio formats supported are MP3 format, WMA format, WAV format.
• We mainly used audio processing libraries like librosa and ffmpeg which are available in python.
• We used Python Flask as the web application framework which connects to the CNN model to give us the results of our query.

Dataset:

• We used the famous GTZAN dataset for our web application. The dataset consists of 1000 audio tracks, each 30 seconds long. It contains 10 genres, each genre consisting of 100 sound clips. All the tracks are 22050 Hz Mono 16-bit audio files in .au format.
• Currently the genres provided in this dataset are:
  • Blues
  • Classical
  • Country
  • Disco
  • HipHop
  • Jazz
  • Metal
  • Pop
  • Reggae
  • Rock
• This application supports various audio formats which makes it more accessible for various users. The more popular examples of audio formats supported are MP3 format, WMA format, WAV format.
• We mainly used audio processing libraries like librosa and ffmpeg which are available in python.
• We used Python Flask as the web application framework which connects to the CNN model to give us the results of our query.

Preprocessing the Data:

• We first split a song into multiple songs using overlapping windows.
• Then, we convert the list of songs to a np array of mel spectrograms.
• We read the data and feed it into our CNN model. We improve the model by using global pooling and MLP (Multilayer perceptron).
• Data is trained and we get val loss = 0.964 and val accuracy = 0.722. The accuracy and loss graphs are attached henceforth.

• Confusion Matrix:

• We can observe that the genres most likely to be mislabeled are - jazz, country, blues & rock.
• The pairs most likely to be confused as each other are - jazz-country, country-rock, blues-country, metal-rock and blues-rock.
• From the majority voting system, we get an average accuracy of 82%.

Tech Stack used:

• Python 3.7.6
• Tensorflow 1.14.0
• Keras 2.3.1
• Librosa 0.6.3
• Ffmpeg 1.4
• Numpy 1.18.1
• Pandas 1.0.1
• Matplotlib 3.1.3
• Scikit-learn 0.22.1
• Flask 1.1.1
• JavaScript
• HTML 5
• CSS 3
• BOOTSTRAP 4

Github link: https://github.com/vedant-manelkar/Genre-Identifier

Snapshots from the website

Atfal Mafkoda: Missing Children

2023

Atfal Mafkoda or "missing children" in Arabic, is an organization based in Cairo, Egypt working to reunite missing children, adults, and seniors with their families and loved ones. Initially starting as a Facebook page, it has grown to nearly two million followers and helped over 3,000 missing people. This initiative was also chosen by Facebook as the most influential initiative for the second time in 2020. They are now interested in applying age progression forensics to pictures of the missing children, since some of them have been missing for several years.

Since children's appearances change rapidly over the years, especially if they go through challenging circumstances like homelessness, it makes it incredibly tough to predict what they look like at present after an extended period of time. To aid in the search of these individuals, our team is planning to create an advanced ML-based facial aging software to expedite the process of predicting what they look like after they get older. This will help reunite them with their families through the help of volunteers sooner.

Generate diverse versions of a face using Generative Adversarial Networks (GANS), integrating factors like homelessness and age into a missing person's image and also incorporate transfer learning to integrate familial facial traits into the generated images for enhanced accuracy. Develop a pipeline for registering these modified faces in a specific format into a database and integrate it with the previous work done. If time permits, we will also explore methods for matching names from the database with Facebook posts using Language Model-based techniques and implement a system for detecting faces within the database.

This is an ongoing project, we are currently working with the clients to figure out the deployment and also test some of our work.

Link to the github repository: https://github.com/BU-Spark/ml-atfal-mafkoda-missing-children/

ISTE Membership Portal Development

2021-22

My team and I led the development of Indian Society of Technical Educations's (ISTE) membership portal. The portal was inaugurated by the former Chief Minister of Maharashtra state, the honorable Mr. Uddhav Thackeray at the 50th ISTE Faculty Convention. My team and I were also personally acknowledged and rewarded for exceptional contributions for this project by the CM.

Here is an article about the event and our project: Article

The portal attracted an impressive 100,000 student members, 5,000 life members, and over 100 institute members during 2021-22 through the website, a 150% increase from the previous year

Tech Stack: React, MongoDB, Express, NodeJS, AWS S3, Docker, and Linux.

We successfully migrated legacy software with MSSQL to MongoDB, enabling efficient management of over 140,000 life members and 1 million student members from 3,000 institutes and organizations across India.

We implemented a range of features including membership registration, report generation, downloading membership certificates and cards, bulk email capabilities, a user management system, integration with an online payment gateway, and receipt generation.

The website is live on this link

The following are some images from the development of the website as well as the inauguration & articles written about it:

Personalized Language tutor using GPT-3

2023

I had developed this project for the AI4ALL '23 batch during my summer internship. My goal was to introduce the world of AI and NLP to the young women in high schools.

The Personalized Language Tutor project leverages the power of OpenAI's GPT-3.5 model to create an interactive and dynamic language learning experience. The project is structured to guide learners through translating sentences, providing detailed explanations, and engaging in conversational practice, all powered by the capabilities of GPT-3.5.

Features:

1. Translation Module:
• Learn to translate sentences from one language to another using GPT-3.5.
• Explore the use of GPT-3 model API calls to automatically detect the source language without explicit user input.
2. Explanation Module:
• Gain insights into sentence structures and meanings through detailed explanations.
• Encourage learners to actively participate by independently discovering methods for utilizing GPT-3 API calls.
3. Practice Conversation Module:
• Engage in conversational practice with the GPT-3 powered chatbot.
• Utilize few-shot learning techniques, comparing performance 'before' and 'after' to highlight the effectiveness of this approach.

The Personalized Language Tutor project not only introduces learners to the capabilities of GPT-3.5 but also guides them through the process of building a practical language learning tool. The combination of structured modules, interactive tasks, and advanced techniques like few-shot learning ensures a comprehensive and engaging learning experience.

The code implementation is available in the Jupyter Notebook at my Github Repository, guiding learners through the development of each function and the integration of the complete language tutor chatbot.

Originality Score: Is a document written by AI?

2023

Project Report

Project Repo on Github: https://github.com/smritae01/CS640-Originality-Score-Project

The advent of Large Language Models and Neural Networks to generate human-like texts and responses has made a lasting impact on how AI will play a considerable role in automating several corporate and academic environments. Our project intends to evaluate the measure of how original a given text is and possibly classify it as AI or Human-written. Currently, ChatGPT [OpenAI, 2022] has the best performance among the tools in the market in terms of generating coherent and persuasive answers. This is the reason it has one of the fastest growing user bases. While it can be used as a tool to assist in all kinds of work, there has been a rapid increase in its misuse. This has made it imperative to include some AI-generated results detectors to counter the growing misuse of the technology. The primary goal of this project is to use DetectGPT, GPTZero, and OpenAI to create a model that formulates an Originality Score that is a cumulative result of the above models on the selected text datasets (obtained from a combination of human-written and ChatGPT-generated texts), thus enabling us to form better conclusions as to how similar the texts generated by the AI language model are to actual human ones.

Our pipeline looked like:

Our project centers on developing a model that can produce the originality scores of texts, including the outputs of Large Language Models (LLMs) by combining AI detector results with LLM outputs as features. We shall use various detectors, including DetectGPT, GPTZero, and OpenAI to produce verdicts and measures that contribute to a specific verdict (Human or AI generated). The dataset used was the GPT-wiki-intro Dataset Features extracted from the dataset were : Average Perplexity, Burstiness, Completely Generated Probability, z-score and a range of accuracy outputs from the Open AI Classifier. The tasks defined for the project were as follows:

• Classification task - using a Multi-layer Perceptron model classifying data into human-generated or machine-generated.
• Weight Generation Task - to generate weights for each classifier to be learned by the model.
• Originality Score Generation - the final desired outcome, a number between 0 to 1 indicating the chance a text was generated by an AI model.

Sudoku Solver using Python and Pygame

2020

This project's core functionality involves employing a Sudoku-solving algorithm, a backtracking algorithm, to systematically fill in the puzzle's grid while adhering to the rules of the game.

As someone who enjoys solving Sudoku puzzles a lot, I made this program keeping simplicity in mind and with the goal of learning about the backtracking algorithm.

Key components of the project include designing an intuitive GUI for user interaction, implementing the Sudoku-solving algorithm, and integrating these elements seamlessly using Python and Pygame.

Link to the github repo for the project: https://github.com/notgitika/sudoku-solver

Line Following Bot using Webots

2021

In this project, a robot equipped with sensors and actuators is programmed to autonomously follow a predefined path marked by a line. Webots, a powerful robot simulation software, serves as the virtual environment for testing and refining the robot's capabilities.

Key components of the project include sensor integration, algorithm development for line detection and following, and the implementation of control strategies to ensure precise and efficient navigation.

Link to the github repo for the project: https://github.com/notgitika/line-following-bot

Video demo of the simulation:

Battle of the Neighborhoods: IBM Data Science Professional Certificate Capstone Project

2021

Introduction: The average person moves approximately 6 times in their lifetime. Almost everyone have experienced a period of searching for a new house. There are many different options and sources to search for, it's almost impossible for us to look one by one. Knowing that, how can we use data science to help us to find our new home?

Since our job offer it is for a specific address, we have selected a few close neighborhoods to live in. Although we have selected these neighborhoods, still there are a lot of apartments for rent. We have a few exigences regarding the nearby venues, and that will be taken into account to choose our home.

Based on definition of our problem, factors that will influence our decision are:

• the apartment must have at least 2 bedrooms
• the monthly value of the rent cannot exceed R$ 1.700,00
• it has to be at least 5km from workplace
• it has to be near at least 1km from a gym and a market
• social and economic characteristics of the neighborhod based on the available data

We will use our techniques in data science to generate the most promissing apartments for us to live in. Economic data and social characteristics of the neighborhoods will be considered in the proccess.

This project is targeted to everyone that might be interested to know more about rent prices of apartments, statistics of the neighborhoods or just want to move to São Paulo, Brazil.

Methodology: In this project we have to find an apartment that fit our needs. In first step we have collected the required data: rental properties data, type (category) of every venue in a 1km radius from each neighborhood. We also have filtered every apartment in a 5km radius from our workplace. In the second step we will be performing a exploratory data analysis in order to extract information about economic and social data of every neighborhood and perform a k-means to cluster the neighborhoods. We will calculate the distance through every selected venue (gyms and markets) to every apartment, in order to find out the nearest venue of each category to every apartment. Finally we will filter our dataset to fit our basic needs regarded as value of the rent, distance to the specific venues and number of bedrooms.

Code & Findings are on my Github: https://github.com/notgitika/Coursera_Capstone/tree/master/Week%205

Visualizations:

A map with the apartments that fit our profile and the nearest venues of each with a circle radius every 1km.

A heatmap to see how our apartments are distributed in the csv file.

Results: Our analysis show that although there are a great number of apartments (3.000) in our original dataset, just a few (17) fits our profile. Our first filter with the aim of get only apartments in a 5km radius for our workplace did not filter a lot of properties (100) and that's due we already had filtered the neighborhoods to scrape the data. The heatmpap show that there are clearly concentration of the apartments in the neighboorhods that we early filtered, resulting in a little bias in our results. To next projects we could consider get all neighborhoods in a 5km radius from our starting point. After we grouped the economic and social data of the neighborhoods, we clustered the neighborhoods into 5 groups. Apparently, the neighborhoods are very similar to each other, resulting in a high concentration in one of the groups (~64% of the neighborhoods). To find our potencial apartments, we applied the final filters regarded as number of bedrooms, minimum distance from specific venues and maximum monthly rent value. The filter were very effective, filtering approximately 99,4% of the results. To generate our final dataset, we grouped every apartment data with the information of the nearest venues of interest and the economical and social data of the neighborhood. This provided us with a broad view of the characteristics of each apartment and its surroundings.

Lorem Ipsum

Self-Supervised Learning for Breast Cancer Image Analysis

2022

My team and I worked on this Machine Learning research project as a part of my undergraduate major project under the Head of Department of Information Technology at KJSIT, Dr. Radhika Kotecha.

We published it in Taylor & Francis as a chapter in the book 'Explainable AI in Healthcare' on July 2023. I was the first author of this research paper.

Link to the publication on Taylor & Francis

Link to the github repo:https://github.com/vedant-manelkar/Project_LY_Group_10

Snapshot from Taylor & Francis publication page

Abstract:

Breast cancer, being among the most widespread terminal illnesses, has the potential to claim the lives of thousands of people each year. Machine learning approaches have shown promise in the biomedical field because of recent advances. However, real-time medical data are hard to obtain, and the number of samples is generally small or unannotated for traditional machine learning approaches to make accurate predictions. The proposed approach uses self-supervised learning to overcome this challenge. Since this model works with unlabeled data and can be downstream from any standard dataset, the proposed approach does not need large labeled datasets. The open-source datasets Mammographic Image Analysis Society (MIAS), Digital Database for Screening Mammography (DDSM), and INbreast are used, and samples are arbitrarily selected from them to make a smaller randomized dataset. An accuracy of 96.7% is achieved using bring your own latent (BYOL) as a pretext task on Resnet18 architecture. This chapter demonstrates that the proposed approach surpasses the contemporary state-of-the-art supervised learning methods for breast cancer detection without compromising accuracy. For deep learning models to be used in crucial fields, they should be reliable, and the output should be made in an interpretable manner to understand the reasoning behind a decision made by the model. For this reason, this chapter further demonstrates explainability for deep learning architectures through the use of saliency maps. This approach has the potential to avoid misdiagnosis and assist medical practitioners in turn, saving innumerable precious lives.

KJSIT Students' Council Website Development

2021

I led the creation of the website while I was the president of the Students' Council at my undergraduate institution.

The tech stack used was:

• HTML
• CSS
• PHP
• Vanilla JavaScript

Here is a link to the github repository (private): https://github.com/notgitika/sc-kjsieit-website

Some screenshots from the website:

Diffusion Adapters for Prior-Preserving Stylization

2023

This is an ongoing project under the Image and Video Computing lab at Boston University with co-authors: Dina Bashkirova, Jiayi Shen, Arijit Ray, Dr. Ranjay Krishna, and Dr. Kate Saenko.

Abstract of the research project: State-of-the-art generative models achieve remarkable expressivity and realism largely due to their rich prior about the visual domain and language-based conditioning. Most applications of image generation require personalization, i.e. ability to adapt to a specific concept or style, with only a few ground truth examples. Existing diffusion-based concept learning methods either require finetuning the entire generator for a single concept which is largely impractical, or propose adaptation in the language space, which results in inaccurate concept representation. In this paper, we propose this novel method, a simple yet effective adaptor-based approach for concept learning. We show that our method achieves accurate concept representation on par with the methods that finetune the entire generator while preserving the prior of the original model, is significantly more accurate than the language-based adaptors, and allows composition of separately learned concepts during inference. We illustrate capabilities of our method in learning a specific object or style, and show that simple prompt-based approach is efficient for disentangling style from other factors in a few and even a single example.

A few examples of the inference outputs:

Cradles 2 Prison: Analyzing Boston's Juvenile Justice Data

2023

The Cradle-to-Prison Pipeline Research Team at Northeastern University conducted an expansive survey, led by Law Professor and Principal Investigator Stephanie Roberts Hartung and Erin Stewart '21, a Skadden Fellow at Citizens for Juvenile Justice, of incarcerated people in Massachusetts.

My team and I worked with this survey data as well as some other prison data to uncover some findings.

The survey addresses critical data gaps in data relating to childhood experiences and systems involvement along the pipeline. The survey was designed to test the hypothesis that children of color are disproportionately impacted by an array of systems—child welfare, public education, mental health, school discipline, and juvenile justice—which operate to increase the likelihood of adult incarceration. The collection of data through this survey is intended to support advocacy efforts and inform policy decisions to help dismantle the cradle-to-prison pipeline and address mass incarceration in meaningful ways.

Link to the github repository: https://github.com/BU-Spark/ml-cfjj-c2p/tree/dev

The goal of this project was to further analyze the extracted data and to answer the client's key questions, and to help clean and compile additional data on school-based arrests.

Below is a poster attached with our findings from the data:

Below are some other interesting visualizations: