I Measured respiration and heart rate through array of pressure sensors and Deep Learning (CNN, RNN-LSTM) algorithms. [Filed Patent Ref No. 201731038573]. 96% accuracy was achieved. The project had three stages 1) preprocessing with analog filters, 2) Digital Signal Processing, 3) Machine Learning model there after data is uploaded to the cloud for further processing.

Collected audio using microphone and Qualcomm’s Snapdragon processor enabled custom built board running Android 6.0. Audio Cleaning, Feature Extraction (more than 700 features) in Temporal, Spectral, Energy, Harmonic, and Cepstral domain, Feature Selection for faster training and better accuracy and ML models (CNN, LSTM, RF, SVM) are designed to classify snoring and other sound and their impact on sleep quality. I used Google Audio dataset to train the model.

Being as research scholar at IIT Kharagpur and working in a project. I used CNN, Inception, LSTM, CNN-LSTM models to estimate Sleep Apnea and Sleep Stages using the respiration signal only. The training data was obtained from MESA, National Sleep Research Resource, USA. I did data cleaning and balancing thereafter feature engineering and selection was done before applying ML model. Achieved 82.7% accuracy in Sleep Apnea which is higher than any known methods applied on respiration signal in MESA Dataset and achieved more than 80% accuracy in sleep stage classification which can be considered as good accuracy.

I have designed my own Automated Emailing software for the marketing. It uses AWS's SES, SNS, SQS, Lambda, DynamoDB and PyQt5. Thereafter, I did web scrapping to collect emails and mobile numbers for more than 10 lakhs people across the globe. Later, I had to drop all those data because of the privacy policy. This automated emailing software could result in huge saving during marketing campaign.

I had performed feature selection and data balancing on Santander Dataset to achieve better accuracy (>96%) than 1st ranker at Kaggle (82.907%). Applied two methods 1) Constant, Quasi-Constant, Duplicate, and Correlated feature removal and 2) Principal Component Analysis, PCA and LDA. In both cases, I achieved better accuracy. Feature Selection helped me to reduce feature dimension from 370 to less than 100 which increases the accuracy and boosted training time. Tried feature selection with ROC_AUC, Univariate ANOVA test, Fisher Score, Chi2 (χ2), Step Forward and Backward, L1 and L2 Regularization, and Recursive Feature elimination.

Used IMDB movies poster image dataset and CNN for Multi-label Image Prediction. The movie poster classified in multiple genre like Drama, Romance, Music, Comedy etc. Learning curve is plotted to understand how model is performing. In another project, I used LSTM for a week ahead House Hold Electricity Consumption prediction. Exploratory Data Analysis (EDA) was done to understand data graphically, it helped me to design algorithms. RMSE was less than the standard deviation of true power consumption.

Class imbalence was major problem in Credit Card Fraud detection project. I used data balancing techniques to improve the accuracy up to 92% using 2 layers of CNN with Dropout and Batch Normalization layers.

I have used WISDM_v1 dataset for the HAR detection. The Dataset was highly unstructured, I had to clean and impute it before start working on it. The EDA helped me to decide which algorithm might work on this. Since this is time series data both RNN and CNN was the best fit. I had used 2D CNN to detect HAR by using only accelerometer data with more than 90% accyracy. In a another project, where malaria parasite affected blood cell was given, I used Image Processing to detect Malaria with more than 92% accuracy.

A Real-time Sentiment Analysis is presentd in this project by using NLTK and Speech Recognition techniques. The phone call voice was automatically transcripted into the long sentences using Google's speech to text api in real-time. Thereafter, sentiment analysis was applied on each sentences.

A Spam Text Classification using spaCy is presented in this project. I did Data Exploratory Analysis (EDA) to understand how total number of text characters are related with SPAM or HAM messages. It was found that text messages with larger number of characters are mostly SPAM messages. Thereafter, I did TF-IDF to convert text messages into numerical values after cleaning the raw text data. TF-IDF is tranined and tested with SVM classifier and then I had also tested it with some real messages received in my mobile phone.

I have experience in feature extraction from Text Files and text extraction from PDF files. Mostly, I have used spaCy to do text processing and written some custom rules as well. Custom rules were inserted in pipeline processing of spaCy for correct working of whole algorithm. I have used Py2PDF for text data extraction from the PDF files.

I have shown a real-time graph of sentiments of twitter users on Donald Trump and Elizabeth Warren who are running for US Presidential Election 2020. There are many steps involved in this process 1. Twitter data extraction 2. Feature extractiong and text cleaning 3. Sentiment analysis for each twitts separately for Donald Trump and Elizabeth Warren 4. Real-time dynamic plot using matplotlib.