My Projects
Hackathon 1 - Voice based food ordering system
The goal of this hackathon was to create and deploy a food ordering application, using speech recognition AI techniques. It was exciting to train a model using features extracted from voice samples and to integrate a machine learning classifier that could identify the correct food items and servings.
It provided an excellent hands-on experience in coding an end-to-end solution, starting from pre-processing the data, to training and testing the model and finally deploying the model. The final user interface accurately identified the food item from a voice command and confirmed the order as a response.
Hackathon 2 - Alexa Chatbot
The goal of this hackathon was to build a conversational bot to interact with the user and achieve the desired outcome, such as booking a doctor’s appointment/ providing a movie recommendation/ assisting in the purchase of a suitable mobile phone, based on the user’s requirements.
Two approaches were used in achieving the task - (i) Amazon Alexa API based bot and (ii) Python coding based bot. It was very interesting and helped to strengthen my understanding and skills learnt in Natural Language Processing (NLP) and Dialog Management. It provided me with an in-depth experience in utterance usage, intent classification, entity design and conversational session creation.
Hackathon 3 - Anti Face Spoofing
In this Hackathon, we implemented AI based Face and Expression recognition to build an anti-face spoofing application. The goal was to unlock a mobile application based on the accurate identification of a user’s face and expression, as a safety feature.
Firstly, a Siamese Network based representation was obtained from a large dataset of face images and used for detecting Face Similarity. Then we built a small dataset of our team members’ face images and expressions and applied a deep learning based CNN model to perform Face and Expression Recognition. It was an amazing experience to unlock the mobile app with our Face and Expression Recognition model, at the end of the project.
Mini Hackathon 1 - Literacy Rate Prediction
The goal of this Mini-Hackathon was to predict whether the literacy rate was high/ medium/ low in the different districts of India. A real world government released dataset consisting of district wise demographics, enrollments, and teacher indicator data was used for this project.
Extensive data-preprocessing was required to clean the data, remove missing data and values, transform the data by normalization, select uncorrelated features and integrate the data into a single dataframe. After the data was pre-processed, a machine learning classifier was trained to obtain a model that could accurately predict the literacy rate (high/ medium/ low). Through this work, we were able to gain excellent experience in all aspects of data pre-processing, model building and prediction.
Mini Hackathon 2 - Author Identification
The goal of this Mini-Hackathon was to identify the unique writing styles of different authors using a machine learning based approach. A manual, stylometric analysis using hand-crafted features was also used for comparison. This project helped me to understand the usefulness of machine learning based text classification for automated author prediction tasks.
A machine learning model was trained on textual data with corresponding author labels, from the Gutenburg database. Bag of Words and Word2Vec based approaches were used for feature representation and analysis. Text classification and author prediction were achieved with >90% accuracy.
Mini Hackathon 3 - Prediction of Client Subscription
The goal of this Mini-Hackathon was to predict if the client would subscribe to a bank term deposit. This is a machine learning based classification task that helped to gain insights on the various factors determining the successful subscription by a client.
This project’s data consisted of details from phone-based marketing campaigns for a bank term deposit, alongwith the resultant label of a successful or unsuccessful subscription. PCA was applied to reduce the data redundancy, followed by a Decision tree classifier to accurately predict client subscription. The evaluation of the predictions was performed with appropriate performance metrics.
Mini Hackathon 4 - Speech classification
The objective of this Mini-Hackathon was to identify and classify simple vocal utterances such as ‘yes’ or ‘no’. This was useful in learning feature extraction from audio files, representing them in a form which can be used as input for a machine learning classification task and visualizing the audio data.
>2600 audio samples were used to extract useful feature representations in the form of Mel Frequency Cepstral Coefficients (MFCC). A machine learning classifier was trained on the MFCC features and the model was accurately able to classify the test audio samples as ‘yes’ or ‘no’
Mini Hackathon 5 - Classification of fashion clothing
The goal of this Mini-Hackathon was to classify fashion clothing by building a machine learning model and improving its performance in recognizing fashion clothing images. This work has practical applications in e-commerce and online advertising, such as identifying similar clothing from a large set of images, automatic inventory assessment in clothing stores, automatic segmentation of clothing for fashion trend prediction.
The dataset used for this project was the Fashion MNIST dataset that consists of 7000 images each, for 10 clothing or accessories categories such as shirts, trousers, bags, dresses, shoes etc. Firstly, features were extracted from the raw images or PCA applied images. A multi-layer perceptron was trained on the features and model performance was evaluated. The model performance was improved using different hyperparameter values, regularizers etc.
Mini Hackathon 6 - Image processing and transformation
The objective of this Mini-Hackathon was to classify images of dogs and cats, after image pre-processing and transformation. This was useful in learning image based pre-processing techniques and achieving image classification using neural networks.
A Pytorch framework was used to build and train a Convolutional Neural Network (CNN) model on a dataset of >22000 images of cats and dogs. Importantly, image classification was achieved with the CNN model and an accuracy of >90% was obtained.
Mini Hackathon 7 - Retail sales forecasting
The goal of this Mini-Hackathon was to obtain the sales forecast for a big retail store. Such a forecasting model would be very useful for the retailer in planning the budgets and investments and minimizing revenue losses.
This was presented to us as a Kaggle competition. The data consisted of a real world dataset that included the monthly sales of the store from the past 5 years. We completed the competition successfully by building a time series based machine learning model that accurately forecasted the monthly and product wise sales for the retail store.
Mini Hackathon 8 - Clustering of academic groups based on research topics
The goal of this Mini-Hackathon was to develop a clustering (unsupervised learning) based tool to identify academic groups working in specific research areas. It was a novel methodology that helped to find the most relevant academic labs for graduate research students, based on their research interests and goals.
Using a large, real-world dataset of biomedical research topics, abstracts, research investigators and their funding records, we performed NLP and k-Means Clustering to obtain research area based academic group clusters. This exploratory data mining approach helped us to identify research groups working on similar research topics.
Mini Hackathon 9 - Identification of plant disease from leaf images
The goal of this Mini-Hackathon was to identify the plant disease from leaf images using deep learning techniques. Such a tool has immense value as it can aid the faster identification and treatment of crop diseases and prevent food and economic losses.
Over 4500 leaf images and their disease labels from the Plant Village dataset were used for this project. Firstly, a denoising autoencoder was built and applied to denoise the images to which noise had been added (to resemble real-world noisy image data). Further, the leaf images were classified to detect the disease, using a computer vision CNN algorithm. This provided us with an end-to-end experience in solving an important real-world problem.
