Hi, I’m Dheeraj Manchandia
a Developer. Product Manager. Data Scientist. AI/ML Engineer.

Featured - Product Management Empurple Enablers

Improving Accessibility for Color-blind users while using Microsoft Teams

Winning team among 10 teams working on similar Product Management projects, awarded by stakeholders from Microsoft.

Improving the Accessibility of Microsoft Products

Ask:
Microsoft has a strong stance on accessibility and inclusivity- choose one Microsoft product and make it more accessible to users with disabilities. Focusing on a core user group, develop accessibility features for a product in Microsoft’s product portfolio. Explore different areas and learn about a group of users that often get left behind.
Your chosen user group must be currently underserved by the offerings in your chosen area. The features that you build should directly address the real problems currently experienced by your users. The solutions must fit within the Microsoft ecosystem, the business and product priorities, and the industry space.
Put simply, help make Microsoft products more inclusive

We started the project by finalizing the Microsoft Product and the user group that we wanted to focus on.
So, we decided to choose Microsoft Teams, and the underserved community of color-blind users

Why Teams and Why Color Blindness?
Microsoft Teams is a popular collaboration platform used by millions worldwide. However, there are still opportunities to enhance accessibility, particularly for those with vision impairments like color blindness.

Colorblind individuals have difficulty distinguishing between certain colors, and this can affect their experience with software and applications that rely heavily on color-coded information or interfaces.

Around 10% of Teams' 280 million users, or 28 million people, are colorblind. The problems they face include lack of customization, low contrast, inaccessible data visualizations, and reliance on color cues.
Now how did we come up with this number, since we didn't have direct access to Microsoft Teams data, this was acheived via the Discovery phase

In this phase, we studied the product i.e. Teams, and our user group i.e. color-blind users in much detail.
Few statistics that we found were:
- The estimated number of people with color blindness is around 300 million worldwide.

- Worldwide Users of Microsoft Teams: 280 Million
- 1 in 8 men and 1 in 200 women are colorblind
- We also found that 72% of Teams users are men and the rest 28% are women.
- So by doing some calculations, we found out:
Estimated Color-Blind Users in Microsoft Teams: 28 million (10%)

The next step was to find out the problems that color-blind users face, now this was a challenge since we didn't have access to Microsoft's historical data and it was even difficult to find color-blind people who use Teams.
So we had to get creative and figure out different ways to correctly identify these problems.

We simulated colorblindness using a browser plugin and tried to use Microsoft Teams as a colorblind user. The below video is a demonstration of this.

We even researched various Facebook pages, Reddit communities, and articles related to color-blindness.
We finally found the major problems:
- No Onboarding Workflow for Accessibility
- Issues With High-Contrast Mode
- Color-coded Information That Can Be Missed
- Issues with Visualization

Once the problems were found, we came up with solutions to fix those problems.
We started this process by using a Framing framework as below:

Now, we have a lot of ideas that we could work on, but we cannot start with all of them at once.
So, the next step for us was prioritization

We considered different models like RICE and MoSCoW, but we went ahead with MoSCoW to scope out features for our MVP

Once our Must Haves were finalized for the MVP, we built a roadmap for their implementation.

Since we didn't have the resources to implement these features we implemented the designs of our solutions to give a vision of our ideas

Finally, to end the project, we defined metrics of success and failures, along with a combination of 4-5 feature indicators for each solution.

To conclude, throughout the project, we followed a structured approach divided into key stages, successfully addressed the core user problems, and aligned our solutions with both the business and user needs.

Featured - Full-Stack | Hackathon | Clustering Combating Food Insecurity

Food insecurity is a pressing issue that affects communities across the globe, and the Atlanta Community Food Bank is no exception. The Food Bank faced several challenges in ensuring frequent, consistent, and convenient delivery of equitable food to partners. Partners had to order a minimum of 6,000 pounds of food for delivery, which sometimes led to missed deliveries or additional costs for renting vehicles. Additionally, partners did not always follow consistent tools, information, and standard practices, and the community struggled to leverage the food items that were being delivered, particularly those who had limited English proficiency.
As participants in the Morgan Stanley CodeToGive Hackathon, we were tasked with developing innovative solutions to these challenges.

Atlanta Community Food Bank faced several challenges in ensuring frequent, consistent, and convenient delivery of equitable food to partners.

1. Partners had to order a minimum of 6,000 pounds of food for delivery, which sometimes led to missed deliveries or additional costs for renting vehicles.

2. Additionally, partners did not always follow consistent tools, information, and standard practices, and the community struggled to leverage the food items that were being delivered

3. Much of the community had limited English proficiency.

1. Our team designed a clustering algorithm that creates clusters of partners near each other, whose orders sum up to more than 6,000 pounds. We also consider the delivery time window for each partner along with the ethnicities that each partner delivers food. This solution addresses the issue of convenient delivery of equitable food, even if the partner's order is less than 6,000 pounds. By doing so, we ensured that partners received the food they needed while optimizing the efficiency of delivery routes.

2. We also recognized the need for consistent tools, information, and standard practices among partners and the community. To address this issue, we developed a learning center that provides tutorials and details on different processes for partners. For the community, we created a learning center that covers nutritional values, recipes, and cooking tips to help people enjoy the food they receive. This approach will encourage the community to better understand the benefits of the food they receive, and how to incorporate it into their daily lives.

3. Finally, we implemented a multilanguage support feature that can be applied to the messaging architecture, inventory management, and other aspects of the system. This feature enables us to better serve people from different ethnicities, regions, and language backgrounds.

Our team designed a clustering algorithm that creates clusters of partners near each other, whose orders sum up to more than 6,000 pounds. We also consider the delivery time window for each partner along with the ethnicities that each partner delivers food. This solution addresses the issue of convenient delivery of equitable food, even if the partner's order is less than 6,000 pounds. By doing so, we ensured that partners received the food they needed while optimizing the efficiency of delivery routes.

- Our algorithm uses KMeans from scikit-learn library for clustering geographic locations based on longitude, latitude, ethnicity, and delivery time window.
- It starts with 2 clusters and iterates up to a maximum of 8 clusters, trying to find the optimal number of clusters using silhouette score as the evaluation metric.
- It sets a minimum threshold for the number of orders per cluster(6000) and checks if any clusters have fewer orders than this threshold. If so, it increases the number of clusters to ensure that each cluster has at least minimum orders.
- The code keeps track of the best silhouette score and the corresponding number of clusters throughout the iterations and finally prints the optimal number of clusters with the highest silhouette score.
- Once the optimal number of clusters is found, it re-runs the KMeans algorithm with that number of clusters and assigns each location to a cluster. It then calculates the total number of orders per cluster and the centroids of each cluster.
- A visual representation of the cluster is shown using TSNE and Matplotlib

We also recognized the need for consistent tools, information, and standard practices among partners and the community. To address this issue, we developed a learning center that provides tutorials and details on different processes for partners. For the community, we created a learning center that covers nutritional values, recipes, and cooking tips to help people enjoy the food they receive. This approach will encourage the community to better understand the benefits of the food they receive, and how to incorporate it into their daily lives.
Finally, we implemented a multilanguage support feature that can be applied to the messaging architecture, inventory management, and other aspects of the system. This feature enables us to better serve people from different ethnicities, regions, and language backgrounds.

- Our algorithm uses the googletrans library to translate text from one language to another.
- It defines a function that takes input text and target language as parameters and uses the Translator class to perform the translation.
- The translated text is displayed on the UI.

- Currently, there is no Maximum quantity mentioned for the orders, we can add that as a parameter based on the amount of food that can be delivered in one cluster and the size of the delivery trucks.
- Forecasting order quantity of the partners, using data analytics.
- Unattended delivery at the partner center using QR codes.

Overall, our Morgan Stanley CodeToGive Hackathon project provides an innovative solution to address the challenges faced by the Atlanta Community Food Bank. By leveraging technology and community outreach, we were able to create an effective and sustainable solution that improves the lives of those facing food insecurity. We believe that by working together and implementing innovative solutions, we can help create a more equitable and sustainable world.

Skills: React.js | Python | Flask | k-means clustering | Project Management

Featured - Data Mining | Machine Learning | Full-Stack TextShield

In today’s digital age, social media platforms have become a common avenue for communication, but unfortunately, they have also become a breeding ground for hate speech and offensive language.

We have proposed an NLP-based approach using the Robustly Optimized BERT Approach (RoBERTa) that can detect not only explicit and direct forms of hate speech but also subtle and implicit forms. By taking into account the broader context in which the language is used, our model can better differentiate between language that is intended to be harmful and language that is not.

By analyzing online text, and data mining algorithms, we aim to identify patterns and features that differentiate between hate/offensive speech and non-hate speech to enable platforms to moderate content effectively and reduce the harm caused by hate speech also ensuring the safety and well-being of individuals using these platforms.

We aggregated data from different data sources and social media sites.

- Handle missing values
- Data cleaning
- Handling contractions
- Data visualization using bar charts, word clouds, etc.

We tried different algorithms mentioned below:
1. LSTM
2. BERT
a. Distil-BERT
b. RoBERTa

- LSTM has an accuracy of 86.36 % on the Validation set and 89.93 % on the testing set.
- Distil BERT model performs much better than the LSTM model with a validation set accuracy of 91.02 % and test set accuracy of 90.04%.
- RoBERTa model performs quite similarly to that of Distil BERT with the validation set accuracy of 92.45 % and test set accuracy of 90.70 %.

We designed a web app that can be used to test and demo our solution.

In this study, we developed a high-quality text corpus of English hate and offensive speech by aggregating data. We conducted an extensive empirical analysis by taking into consideration three models and built our project.

Skills: Data Mining | Data Analysis | Machine Learning | Data Visualization | Full-Stack Development

Featured - Full-Stack | Software Enginnering | Agile TourWizard

Efficient and organized tour management services

TourWizard was initiated to address the increasing demand for efficient and organized tour management services. With the rise in tourism, there was a growing need for a system to streamline the process of organizing tours, from creating itineraries to managing bookings and payments.

The manual process of managing tours is cumbersome and time-consuming. A tourist must visit different applications for different purposes like finalizing the places to visit, searching for accommodation, identifying travel routes, getting those tickets, fixing a schedule, and locking on to that itinerary after several rounds of changes. TourWizard can help to improve the experience for tourists.

This website lets users search for places, cities, accommodations, and travel bookings. The application will generate an itinerary based on the user's selections.

- Allow users to browse the site without logging in
- If the user tries to book, save a place, or create an itinerary then they have to login/signup
- Users can search for cities, accommodations, and flights
- Users can browse their selections on the Itinerary page
- Once users have finalized their tour, they can even book and do the payment for all the selections in their itinerary

Skills: Team Leadership | Agile Methodologies | MongoDB | Full-Stack Development | React.js | Flask | Project Management

Featured - Machine Learning | CNN | Classification Object Detection

A classifier that decides the correct category of a given image

Object detection is a well-explored topic. In this project, the main task is to recognize the object in the images, i.e., to create a classifier that decides the correct category of a given image.

The training set contains 100 categories with labels from 0 to 99.
Each category has 5000 images.
Each image is formatted as 28 × 28 pixels.
There are 500K images in total.
The test set covers the same categories in the training set without any unseen categories. The test set has 100K images.

Grayscale and Downsize
Each image has one object only.

Many classification algorithms can be used, a few of them that I considered were:
• Logistic Regression
• Naiye Bayes
• K-Nearest Neighbors
• Decision Tree
• SVM
• CNN
Now from the above-mentioned algorithms, the best ones for Image classification are KNN and CNN
With KNN, the accuracy (44.48%) was not too good, so I decided to implement the project using CNN

I implemented a CNN model to train the data
I used Keras API to implement it and stack the layers of my deep network
- Staring with a convolution layer, having 64 filters and each filter is 7 x 7
- The next layer is a max pool with a size of 2
- Then few more layers of convolution, adjusting the filters and their sizes to improve the accuracy
- This is followed by a flattened layer, which will take the output of the convolution and vectorize it, so it can pass through a multi-level perceptron or dense layers
- I have also used Dropout
- In the end, the output will have 100 units with activation as SoftMax

After training and fitting the model, I used 10 epochs
I got an accuracy of approximately 52%
To improve the accuracy, various regularization and optimization techniques can be used:
• Parameter sharing
• Weight decay
• Dropout
• Early stopping
• Change of learning rate during learning
• Adding momentum
• Dataset augmentation
• Activation function selection

Skills: Data Classification | Convolutional Neural Networks (CNN) | Machine Learning | Python