Projects - Details
Auto Response Deal
Problem Statement: To generate auto-response samples to assist agents in responding faster & better to incoming customer emails based on the contents of the emails using NLP powered solution. NLP powered solution for 1st level of auto-reply that generates 3 sample responses for the agents to choose from to reply. This solution is set up and integrated into the zendesk environment. The solution has an active learning pipeline packaged within the solution.
Output: Generate top 3 email response for the customer query and assigned to the best agent related to query.
Research paper used in AI-Model.
1. BERT | Source: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
2. OpenAI-GPT2 | Source: Language Models are Unsupervised Multitask Learners
3. Smart Reply: ARS-Email | Source: Smart Reply: Automated Response Suggestion for Email
Audio classification based on Connector clicks
Problem Statement: Develop machine learning solution for detecting the clicks helping engineers to ensure that connections in the vehicles have been made properly in the plants. The end solution will be a user interface that needs to be integrated with the hardware to take real-time decisions.
Data-set : 10 hrs. Audio data for analysis and to build out model for dectection clicks.
Output: Tag and label the audio data. Preprocess and clean the data for modeling. Train a classification model using CNN and RNN. Integrate the model with a user interface. Deploy and host the solution on Google cloud platform and develop the CD pipeline for deployment.
Research paper used in AI-Model.
1. Audio Event Detection | Source: Audio Event Detection Using Deep Neural Networks
2. Deep Learning for Audio Signal Processing | Source: Deep Learning for Audio Signal Processing
Monitor the Person Appearance
Problem Statement: Monitor the person’s attendance. Associate the faces with the existing employee database to identify each employee record the time stamp and date for each employee recognized.
Data-set : 100 hrs. Video footage for analysis and to build out model for monitor of person appearance.
Output: The output of the analysis will be recorded in an csv file, which will record the person’s id and the his/her time stamp and date
Research paper used in AI-Model.
1. Object Detection | Source: SSD: Single Shot MultiBox Detector
2. Face-Recognition | Source: FaceNet: A Unified Embedding for Face Recognition and Clustering
3. Face-Tracking | Source: An Algorithm For Centroid-Based Tracking of Moving Objects
DNA-Splice Gene Prediction
Each DNA read is a sequence of four [C,A,G,T] types of nucleotides and needs to be converted into numerical representations for machine learning.
The domain consists of 60 variables, representing a sequence of DNA bases an additional class Variable. The task is to determine if the middle of the sequence is a splice junction and what is its type: Splice junctions are of two types:
1. exon-intron (EI): represents the end of an exon and the beginning of an intron
2. intron-exon (IE): represents where the intron ends and the next exon, or coding section, begins.
So the class variable contains 3 values:
1. exon-intron (EI) | 2. intron-exon (IE) | 3. No-Junction
Other 4 values corresponding to the 4 possible DNA bases (C, A, G, T)
C : Cytosine | A : Adenine | G : Guanine | T : Thymine
Research paper used in AI-Model.
1. DNA-Splice-Jucntion | Source: DNA-Level Splice Junction Prediction using Deep Recurrent Neural Networks
2. Classification of DNA-Splice | Source: Classification of eukaryotic splice-junction genetic sequences using averaged one-dependence estimators with subsumption resolution
Quality Enhancement of Images Using GAN
Problem of this project is that we have to build a model which makes it possible to generate HR-Images based on their LR-Images
A deep-learning-based solution for the construction of a super resolution images. Trained the model on 200k LR-Images and the condition to be enhanced of LR-Images.
Research paper used in AI-Model.
1. SR-GAN | Source: Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
2. Condition SR-GAN | Source: Class-Conditional Superresolution with GANs
Data Warehouse System
The problem of this project is that we have to design a scalable data warehouse system that contains various categories and classified image in that categories with no duplication and various different operation. Trained the model on more than 100k images for our data warehouse system and perform different operation on these images. Integrate the whole system in the pipeline and create a GUI and API for the user interface. The accuracy of the system is 99.38 %.
Research paper used in AI-Model.
1. VGG-16 Architecture | Source: Very Deep Convolutional Networks for Large-Scale Image Recognition
2.Hashing Techniques | Source: Selecting a Hashing Algorithm
2. Simialrity Algorithm | Source: An Online Algorithm for Large Scale Image Similarity Learning
Recommender System
The problem of this project is that we have to design a recommender system for male & female based on their dressing style and face shape etc. For this project, we have used the detection and classification model. Trained the model on the huge number of label images for our detection and classification model and perform different operation on these detected images. Integrate the whole system in the pipeline and create a GUI and API for the user interface.
Research paper used in AI-Model.
1. Object Detection | Source: You Only Look Once: Unified, Real-Time Object Detection
2. Classification | Source: Rethinking the Inception Architecture for Computer Vision
3. Clustering | Source: Application of k-Means Clustering algorithm for prediction of Students’ Academic Performance
4. Recommender System (using Bi-Directional LSTM) | Source: Long Short-Term Memory Recurrent Neural Network Architectures for Large Scale Acoustic Modeling
Chatbot for Automatic Question Answer
The problem of this project is that we have to build a chatbot which makes it possible to generate responses based on collections of known conversations. A deep-learning-based solution for the construction of a chatbot off-topic behavior and personality. When confronted with off-topic questions, the bot will try to automatically generate a possibly relevant answer from scratch, based only on a pre-trained RNN model.
Course Projects
Face Shape Type Detector of Human Being
Implemented a system that analyses various type of face like square, oblong, oval, heart, round and diamond. Dataset gathered from various internet resources and clean the data according to our requirements. Trained the model on more than 10000 labeled images for face shape type detection. The accuracy of system is 84.56%.
Keyword Extraction and Semantic Tag Prediction [KESTP]
APPLIED SUPERVISED LEARNING METHODS FOR PREDICTING TAGS
• Implemented a system that analyses the tags for a question asked by user into appropriate categories.
• Data set of Stack overflow dump is used to predict the tags according to user questions.
• Applied java sax parser to parse the useful data from data dump.
• Applied different classifiers on title ,body and code to predicts the tags.
• Used M.L algorithms like SVM , Naive Bayes & Unique Feature Extraction to predict tags.
House Price Prediction-Regression Techniques [HPPRT]
APPLIED MACHINE LEARNING ALGORITHM
• Devised a mechanism to predict the price of each house using stacking of various outputs of regression techniques like lasso and ridge regression, XGboost, elastic net.
• This Project is implemented as the part of kaggle competition.
Stock Price Trend Forecasting [SPTF]
APPLIED SUPERVISED LEARNING METHODS FOR STOCK PRICE TREND FORECASTING
• Implemented a system that analyses previous stock data of various
companies, processes time-series data and aims to forecast the trends of stock in near future.
• Data set of Indian stock market is used to predict the next-minute, next day and next week stock trend.
• Applied moving averages,Log normalization techniques for feature generation & selection process.
• Used M.L algorithms like SVM & Neural network to predict trend.
Search Engine for 81 GB English Wikipedia Dump
NFORMATION RETRIEVAL
• Implemented inverted index and calculated TF-IDF score.
• Retrieving the top 10 results for any query search within 0.3-0.6 seconds.
• Supports field search on title, infobox, text, categories.