Description

---

The project "Driver Drowsiness and Emotion Detection" is a machine learning project that aims to identify whether a driver is drowsy or not by analyzing their facial features, as well as detecting their emotions in real-time. The project uses various libraries, including Haar Cascade and Dlib. The project is implemented using a deep neural network architecture that is trained on a dataset of images that include facial features of drivers who are both drowsy and alert, as well as images that show a range of emotions. The network can be trained to identify specific features associated with drowsiness, such as heavy eyelids or a downturned mouth, as well as the characteristic features of different emotions. The project is be divided into two parts: drowsiness detection and emotion detection. For drowsiness detection, the system can use eye tracking technology to detect when a driver's eyes are closed or partially closed for an extended period of time, indicating drowsiness. For emotion detection, the system can analyze facial expressions to determine the driver's emotional state, using techniques such as facial landmark detection and tracking. The project is be developed using Python and various libraries such as OpenCV, Keras, TensorFlow, and NumPy.

---

Description

---

The objective of this project was to compare different machine learning models and select the one that predicts customer churn best using R.

• The dataset includes information on banks and their current and former customers, with 10,127 rows and 21 columns. The primary target for prediction is "Attrition Flag."
• Data cleaning involved handling missing and duplicate values and dropping the last two columns with garbage values. The data was further preprocessed with Principal Component Analysis (PCA) for dimensionality reduction, resulting in retaining 10 principal components that explained 95.713% of the variance.
• Multiple machine learning models were tested. The Random Forest model emerged as the best performer with an accuracy of 91.26% for dataset I and 96.35% for dataset II.

---

Co-created a YouTube channel to curate music of Indian Electronic Dance Music (EDM) genre during undergrad as a passion project.

Description

---

• Published over 180 videos and grew the YouTube channel to 580,000+ subscribers, 13000+ Instagram followers, and 6000+ Soundcloud followers.
• Managed promotional Facebook and Google ad campaigns for over 20 artists.
• Received Silver Play Button from YouTube creator awards team for crossing 100,000 subscribers.
• Developed a bootstrap website ( turbantrapofficial.com) to automate music submission process.
• Designed an AR filter using Spark AR Studio that was shared over 1.7 million times on Instagram and Facebook.
• Deployed YouTube API v3 to fetch estimated revenue using flask so artists can determine individual profit share.

---

The main idea of the project is to show which amongst the K-NN, SVM, and Neural Networks is better classifier in the machine learning set and which performs the best on which type of data. The objective is to predict whether a patient has diabetes or not using these models by factoring in independent variables such as Cholesterol, Smoking Habits, Sex, BMI, Income, Age etc.

Description

---

Models Used: Logistic Regression, K Nearest Neighbors (KNN), Support Vector Classifier (SVC), Naive Bayes, Decision Tree, Random Forest, and Neural Network.

---

Created a YouTube channel to create and publish stickfigure animations of comedian Bill Burr's podcast clips.

Description

---

Bill Burr is an American stand-up comedian who does an ‘audio only’ podcast twice a week. Since I had newly learned how to edit videos using Premiere Pro and After Effects, I wanted to create something entertaining. I figured creating stick figure animations for specific clips from the podcast would perform well on YouTube. Upon obtaining permission to use the audio clips from Bill’s manager and producer, Andrew Themeles, I started uploading content on this channel in September 2017. After uploading only a few videos, the algorithm picked up a few of them, and the channel has been steadily growing ever since. I started this channel to practice my video editing skills with no other expectations, but it later grew into a nice little side hustle that I am proud of. Currently, it has over 160,000 subscribers.

---

Oliveboard is a Banglore-based online exam preparation platform for Banking & Government Exams like UPSC, SSC, Railways, Police, etc.

Description

---

• Analyzed data from various social media channels to track key performance indicators (KPIs) and formulated content strategy with the objective of maximizing reach, engagement and click-through rates (CTRs).
• Created a Python script leveraging Graph API to automate content scheduling, cutting process time by 15%.
• Collaborated with senior management to draft 50+ Meta and Google PPC campaigns to generate leads.
• Performed analysis on marketing data, derived insights, crafted Excel reports to recommend a course of action.

---

 

A Python application deployed using Flask to compare followers.html and following.html files, and figure out the list of people who have unfollowed you.

How do I use this?

---

• Step-1 Go to your profile, click the 3 lines at the top right corner of your screen. Tap on 'Your Activity' option. Hint - It has a clock as an icon. Now scroll down and tap on the 'Download your information' button and request download. It will authenticate you, confirm your email address and send you the data in about 4-6 hours.
• Step-2 Go to your email, you would have received a new email from IG with subject as 'Your Instagram Information'. Download the attached file, and extract it. It will create a folder with this format 'yourusername_date'. Open it, open a subfolder named 'followers_and_following'. You will find a file named followers.html and a file named following.html, these two are important for us.
• Step-3 Go to https://akshaysinngh.com/unfollowers/ (or simply click the Unfollowers button at the top).
• Step-4 Upload followers.html, and then following.html and press the submit button.

---

 

Description

---

• Integrated LangChain with OpenAI’s GPT-3.5. Turbo model to process movie scripts from IMSDB.com
• Engineered prompt to extract a dictionary object containing summary, important keywords, and word counts. Applied TikToken for token measurement and MapReduce for efficient handling of texts exceeding 4,097 tokens.
• Benchmarked model’s summarization capabilities with a ROUGE score of 0.73 against IMDB’s summary.

---

Technologies Used: LangChain, OpenAI, and Python

---

 

Description

---

• Developed and executed an end-to-end data analysis pipeline for a retail store, leveraging AWS S3 for data storage and Snowflake for data warehousing. Utilized Python in Jupyter Notebook for data transformation and cleaning, ensuring high-quality data preparation for analysis.
• Established a robust data integration process, connecting AWS S3 buckets to Snowflake, enabling seamless data extraction and loading into Snowflake tables. Enhanced data accessibility and manipulation by creating a connection between Snowflake and a local Jupyter Notebook environment.
• Designed and implemented an interactive dashboard in Tableau, utilizing Snowpipe to stream data from Snowflake, providing real-time insights and visualizations of retail trends based on comprehensive datasets, including transactions, products, coupons, and demographics.
• Automated the data update process by setting up a cron job in Jupyter Lab, integrated with Amazon SQS, to trigger updates in the Tableau dashboard whenever new data is uploaded to the AWS S3 buckets, ensuring real-time data accuracy and relevance in visual analytics.

---

Technologies Used: Snowflake, AWS, Python, SQL, and Tableau.

---

 

ELI5 is short for “Explain Like I'm 5,” a request for a simple explanation to a complicated research paper in healthcare or medicine.

Why Do We Need This?

---

The project aims to make the insights from medical research papers accessible to all by simplifying complex jargon and dense scientific language. Research papers, often written by and for medical professionals, can be challenging for the general public to understand due to their technical nature. This project extracts abstracts from research papers, translates them into easy-to-understand language, and visualizes any embedded data or conclusions. By doing so, it democratizes access to scientific knowledge in healthcare, allowing a broader audience to grasp and benefit from medical research findings without needing a biostatistics background.

Example

Effect of Weekly Subcutaneous Semaglutide (Ozempic) vs Daily Liraglutide (Victoza and Saxenda) on Body Weight in Adults With Overweight or Obesity Without Diabetes.

---

---

Technologies Used: LangChain, OpenAI, and Python.

---

 

Labelmaster is a Chicago based company, that specializes in helping businesses comply with regulations for the transportation of hazardous materials.
The main objective of our team was to come up with a machine learning model that would optimize Google Ads expenditure, aiming to either reduce costs, or increase conversions, or ideally achieve both.

Description

---

• Leveraged Tableau's dashboards to conduct exploratory data and trend analysis in 6+ years of Google Ads data.
• Forecasted an 18 month bidding price trajectory for all products through ARIMA time-series analysis.
• Implemented PuLP linear programming model with custom constraints to optimize bidding prices for all 40 campaigns, resulting in a reduction of advertisement costs by about 12%.

---

Technologies Used: Python, Machine Learning and Tableau.

 

This is a lighthearted project that playfully examines the distinct experiences and behaviors of the two genders.It explores aspects like swiping tendencies, age preferences, prevalent professions, and the influence of the initial message on a conversation.

Description

---

• Processed unstructured JSON data with 2 million+ rows, anonymized user names to ensure data confidentiality.
• Performed trend analysis using Tableau on user behaviors, swiping patterns, time spent, selection preferences and messaging habits, differentiated by gender. Published the insights on Medium.

---

Technologies Used: JSON, Python and Tableau.

---

 

As a huge UFC fan, I've developed a machine learning model that predicts the outcomes of upcoming or hypothetical UFC matchups.

Description

---

• Extracted 20 years of UFC event data using BeautifulSoup, followed by extensive data cleaning and EDA. Automated the data scraping and cleaning workflow with a CRON job setup.
• Deployed various Machine Learning models (Linear Regression, SVM, Decision Trees, Naïve Bayes and Random Forest) on Sreamlit to predict fight results with 83% accuracy.

---

Technologies Used: JSON, Python, Machine Learning and Streamlit.

---

 

The task was to collect annotated joint positions of a mouse, including the nose, ears, hips, tail base, and tail end, from a set of approximately 125 photos. The goal was to train a model that could predict the mouse's location given the annotated joint positions.

Description

---

• The project was divided into two main sections. In the first part, I collected and annotated the mouse joint positions using the DeepLabCut library and exported the data as a CSV file.
• In the second part, I built a Neural Network to predict the mouse location based on the annotated joint positions. The model was evaluated using test data.

---

 

Description

---

• This project focuses on the task of fake news detection. The aim is to build a classifier that can effectively distinguish between fake news and genuine news articles based on their titles.
• The dataset provided for this project consists of news pairs, each containing a fake news title (title1_en) and a corresponding news title (title2_en). The dataset is labeled with three categories indicating the relationship between the news pairs: agreed, disagreed, or unrelated. The task is to develop a machine learning model that can predict the relationship label for the test data, which does not contain any labels.
• To preprocess the data, stemming techniques were applied to the news titles.
• For the classification task, Random Forest model was chosen. The model was trained on the provided training data, which includes the labeled news pairs, and the performance was evaluated using the validation data.

---