Binil Kuriachan

Business Email Compromise (BEC) is one of the major Phishing mail categories that pertains major threat to the Enterprise Customers. In 2022, BEC led to loss of ~3 Billion dollars in US alone. I have been working on building an efficient solution to detect and prevent BEC attacks originates from Email for Enterprise.

Focus areas include:
* Traffic selection
* Data pipeline for training
* ML system design
* Design and Build Scalable solution
* Deployment
* Real-time… Show more

Business Email Compromise (BEC) is one of the major Phishing mail categories that pertains major threat to the Enterprise Customers. In 2022, BEC led to loss of ~3 Billion dollars in US alone. I have been working on building an efficient solution to detect and prevent BEC attacks originates from Email for Enterprise.

Focus areas include:
* Traffic selection
* Data pipeline for training
* ML system design
* Design and Build Scalable solution
* Deployment
* Real-time Inference Show less

Built and deployed LLM Grader for Web as part of Microsoft Security Copilot

Building AI solutions for Cybersecurity in M365 product Suite
 Phishing Mail Identification & Generation (Attack Simulation Module)
 Predicted Compromise Rate, Phish Susceptibility Score (User level)
 Web Defense System for Web Protection

Crop Classification using multi-spectral Geo-Spatial Data (Satellite/Weather) for Azure FarmBeats platform. Identification of crop types early in the growing season aids in many pre-harvest decisions, including monitoring food security, directing the best use of the landscape, supporting agricultural policy, enhancing supply chain efficiency through crop yield forecasting, and estimating crop water needs for irrigation planning. Therefore, there would be substantial advantages if a model could… Show more

Crop Classification using multi-spectral Geo-Spatial Data (Satellite/Weather) for Azure FarmBeats platform. Identification of crop types early in the growing season aids in many pre-harvest decisions, including monitoring food security, directing the best use of the landscape, supporting agricultural policy, enhancing supply chain efficiency through crop yield forecasting, and estimating crop water needs for irrigation planning. Therefore, there would be substantial advantages if a model could accurately predict crop type early or mid-season. We built a robust Deep Learning Model that can classify the crop types by using temporal patterns emerging from multitemporal satellite images and past year's crop rotation patterns. This model can predict the crop type at the pixel level for a given boundary with latitude/longitude coordinates. CDL data is used in our study to label training samples on satellite images for crop type classification and to track crop rotation patterns from previous years. The greenness and density of the vegetation visible in satellite images are measured using the Normalized Difference Vegetation Index (NDVI), which exhibits different temporal patterns for different crops during a growing season. Crop rotation is used to promote soil health and increase the amount of nutrients available for crops since monocropping, the practice of growing the same crop in the same location for an extended period, gradually depletes the soil quality. To give the model more intelligence, crop rotation signals from prior years are incorporated. A sequence-to-sequence deep learning model is developed utilizing crop rotation signals and temporal inputs to attain an overall accuracy of 90%. The accuracy can be improved by using more temporal data over a growing season. Due to its multimodality, our suggested method can be applied to a variety of scenarios, including early, mid, and late season predictions. Show less

Azure FarmBeats is a business-to-business offering available in Azure Marketplace. It enables aggregation of agriculture data sets across providers. Azure FarmBeats enables you to build artificial intelligence (AI) or machine learning (ML) models based on fused data sets. By using Azure FarmBeats, agriculture businesses can focus on core value-adds instead of the undifferentiated heavy lifting of data engineering. As part of FarmBeats, we built an NDVI forecast model using 60 days of historical… Show more

Azure FarmBeats is a business-to-business offering available in Azure Marketplace. It enables aggregation of agriculture data sets across providers. Azure FarmBeats enables you to build artificial intelligence (AI) or machine learning (ML) models based on fused data sets. By using Azure FarmBeats, agriculture businesses can focus on core value-adds instead of the undifferentiated heavy lifting of data engineering. As part of FarmBeats, we built an NDVI forecast model using 60 days of historical satellite imagery, 30 days of historical weather and 10 days of forecasted weather data. This project demonstrate different capability of Azure FarmBeats and how it can be used effectively for building end-to-end machine learning models for agriculture and food industry. Show less

Data Imputation in geospatial time series data with deep learning techniques.

Developing Smart Search which has the capabilities of Key Word search with reduced need of indexing along with Artificial Intelligence that helps in understanding the semantics of user query and find most relevant information. We have developed indigenous Named Entity Recognition and custom function to identify entities from the search text. Key Word search module of Smart Search pre-process the document corpus and transforms it into distributable matrices comprising of numerical scores for… Show more

Developing Smart Search which has the capabilities of Key Word search with reduced need of indexing along with Artificial Intelligence that helps in understanding the semantics of user query and find most relevant information. We have developed indigenous Named Entity Recognition and custom function to identify entities from the search text. Key Word search module of Smart Search pre-process the document corpus and transforms it into distributable matrices comprising of numerical scores for each words present in each documents. The matrix can be distributed for efficient computation and an index is maintained which explains the word and its column position. Numerical scores can be calculated using many methods such as TF-IDF, count vectorizer, binary vectorizer etc. Once the search text arrives, it is also transformed into a vector of numerical scores corresponding to the number. The relevant documents can be identified by finding the dot product of input vectors and respective vectors from corpus matrix. In this work, various Machine Learning architectures are also experimented on the document corpus to calculate a word embedding for each word based on its context in the documents. Word Embedding represents a word into a vector of numbers where each value can be seen as a representation for different features which explain the word. Embeddings such as Word2Vec and Glove are used and weights are updated from the corpus. The document corpus is also used to learn the proportion of different topics using Latent Dirichlet Allocation algorithm which is one of the popular unsupervised learning method for identifying latent topics from text document. The calculated Word Embedding and LDA are used to understand the semantic similarity of user query and documents. In order to limit the search space, classification models are trained from the dataset. Smart Search encompasses the best capabilities of learning from the data as well as the Key Word search approach.
Show less

The AoS Analytics Platform is developed to help users identify and address the drivers of part shortages across the integrated value chain from program start-up through engineering requirements release, purchase order placement, supplier delivery, and manufacturing build. This project focus on building insights and use predictive analytics capability to help reducing part shortages and efficient planning.

FSAT is an architecture tool which helps users by suggesting the possible future architecture of an application based on the current architecture of the tools. In this project, we created a Deep Learning based NER model with Bidirectional LSTM (Long Short Term Memory) Network backed with Convolutional Neural Networks (character, word and case embedding) for detecting technical keywords from Boeing Propitiatory technical documents and associating it with the respective categories. Model… Show more

FSAT is an architecture tool which helps users by suggesting the possible future architecture of an application based on the current architecture of the tools. In this project, we created a Deep Learning based NER model with Bidirectional LSTM (Long Short Term Memory) Network backed with Convolutional Neural Networks (character, word and case embedding) for detecting technical keywords from Boeing Propitiatory technical documents and associating it with the respective categories. Model performance was evaluated with f1 score.
Tools and Technology: TensorFlow, Python, NLP, CNN, LSTM
Show less

Line Move project focus on identifying different equipment (using camera feed) being used during flight assembly and predicting the possible sequence for better organizing and understanding anomalies.

In this project, two separate models (Deep Learning - CNN and RNN) for Object Detection and Sequence Modeling. In the Object detection task for identifying various equipment being used in the factory line move of flight assembly, we leveraged few of the state-of-the-art architectures like… Show more

Line Move project focus on identifying different equipment (using camera feed) being used during flight assembly and predicting the possible sequence for better organizing and understanding anomalies.

In this project, two separate models (Deep Learning - CNN and RNN) for Object Detection and Sequence Modeling. In the Object detection task for identifying various equipment being used in the factory line move of flight assembly, we leveraged few of the state-of-the-art architectures like YOLO, Fast-RCNN, Mask-RCNN etc and trained it on custom images. After evaluating the performance, Model created with YOLO was selected for deployment. For the Sequence modeling task of identifying equipment sequence, we created a Stacked LSTM network with 200 memory cells and variable length inputs to predict the upcoming equipment requirements.
Technology Used: Python, TensorFlow, Keras, Convolutional Neural Networks, Recurrent Neural Networks, Sequence to Sequence Models, LSTM etc.
Show less

Created a Deep Neural based Encoder-Decoder sequence model for handling multilevel classification at single shot. In this project, we had input text which consists of customer facing issues which usually end up in multiple queues for getting the final fix. Our aim was to identify the sequence of queues without any manual intervention. We could solve it by using a sequence model with Encoder-Decoder architecture as input and output records have multiple length.

Tools and Technologies… Show more

Created a Deep Neural based Encoder-Decoder sequence model for handling multilevel classification at single shot. In this project, we had input text which consists of customer facing issues which usually end up in multiple queues for getting the final fix. Our aim was to identify the sequence of queues without any manual intervention. We could solve it by using a sequence model with Encoder-Decoder architecture as input and output records have multiple length.

Tools and Technologies Used: Python, Tensorflow, Deep Learning, Sequence Models, Encoder - Decoder architecture, LSTM, GRU etc. Show less

In this Work, we build a multi-phase Machine Learning model which predicts the contents of the email and trigger the response template to the user. We build a first level Multinomial Classification model (SVM, Ensemble Models, Deep Learning etc..) which predicts the high level groups and subsequently employed a Topic Model using Latent Dirichlet Allocation (LDA) to figure out the topic distribution. Tools and Technology: NLTK, Scikit-Learn, Python, Bag-Of-Words, Vectorization, Data Mining… Show more

In this Work, we build a multi-phase Machine Learning model which predicts the contents of the email and trigger the response template to the user. We build a first level Multinomial Classification model (SVM, Ensemble Models, Deep Learning etc..) which predicts the high level groups and subsequently employed a Topic Model using Latent Dirichlet Allocation (LDA) to figure out the topic distribution. Tools and Technology: NLTK, Scikit-Learn, Python, Bag-Of-Words, Vectorization, Data Mining, Nearest Neighbors, Document Similarity, Machine Learning , Support Vector Machines, Logistic Regression With Regularization, Random Forest, Word Embedding, Semantic Analysis, Deep Learning. Show less

In this project, we build a Knowledge Base Resolution prediction by using Vectorization along with Semantic Document Similarity on previous ‘Incident Records’ as well as ‘Knowledge Articles’. For improving the accuracy of prediction, incident and knowledge records are filtered by the configuration item using supervised Machine Learning Classification Models (Deep Learning, CNN, Word Embedding, SVM, Ensemble Models, Logistic Regression, Naive Bayes). Tools and Technology: NLTK, Scikit-Learn… Show more

In this project, we build a Knowledge Base Resolution prediction by using Vectorization along with Semantic Document Similarity on previous ‘Incident Records’ as well as ‘Knowledge Articles’. For improving the accuracy of prediction, incident and knowledge records are filtered by the configuration item using supervised Machine Learning Classification Models (Deep Learning, CNN, Word Embedding, SVM, Ensemble Models, Logistic Regression, Naive Bayes). Tools and Technology: NLTK, Scikit-Learn, Python, Bag-Of-Words, Vectorization, Data Mining, Nearest Neighbors, Document Similarity, Machine Learning , Support Vector Machines, Logistic Regression With Regularization, Random Forest, Word Embedding, Semantic Analysis, Deep Learning, Keras, TensoFlow. Show less

With the increased popularity of the smart mobile devices, mobile applications (a.k.a apps) have become essential. While the app developers face an extensive challenge to improve user satisfaction by exploiting the valuable feedback, the app users are overloaded with way too many apps. Extracting the
valuable features from apps and mining the associated sentiments is of utmost importance for the app developers. Similarly, from the user perspective, the key preferences should be identified… Show more

With the increased popularity of the smart mobile devices, mobile applications (a.k.a apps) have become essential. While the app developers face an extensive challenge to improve user satisfaction by exploiting the valuable feedback, the app users are overloaded with way too many apps. Extracting the
valuable features from apps and mining the associated sentiments is of utmost importance for the app developers. Similarly, from the user perspective, the key preferences should be identified. This work deals with profiling users and apps using a novel LDA based aspect identification technique. Polarity aggregation technique is used to tag the weak features of the apps the developers should concentrate on. The proposed technique has been experimented on an Android review dataset to validate the efficacy compared to state-of-the-art algorithms. Experimental findings suggest superiority and applicability of our model in practical scenarios.

Tools and Technologies: Latent Dirichlet Allocation (LDA), Semantic Analysis, Python, R, Natural Language Processing, NLTK, Scikit-Learn, Tableau, Machine Learning Show less

This project is to identify the respective sentiment distribution across multiple topics from the free form text (feedback/survey/Collaboration tools, Chat data etc.). We built a LDA based topic modeling clubbed with Dictionary as well as ML based sentiment Analysis to aggregate the respective polarity across each topics. The results are used for management decision and fine tuning the respective work areas. For Aspect Identification, we developed Latent Dirichlet Allocation (LDA) model and for… Show more

This project is to identify the respective sentiment distribution across multiple topics from the free form text (feedback/survey/Collaboration tools, Chat data etc.). We built a LDA based topic modeling clubbed with Dictionary as well as ML based sentiment Analysis to aggregate the respective polarity across each topics. The results are used for management decision and fine tuning the respective work areas. For Aspect Identification, we developed Latent Dirichlet Allocation (LDA) model and for ML based Sentiment Analysis, models using SVM, CNN+ RNN (LSTM), Logistic Regression with L2 Regularization, Naïve Bayes and Random Forest are evaluated. TextBlob is used for Dictionary based Sentiment Analysis
Tools and Technology: NLTK, Scikit-Learn, Keras, TensorFlow, Python, Bag-Of-Words, Word Embedding, Word2Vec, Vectorization, Machine Learning , Support Vector Machines, Logistic Regression With Regularization, Random Forest, Word Embedding, Semantic Analysis, Deep Learning, Gensim. Show less

Data from machines were given and we created a regression model to predict the range and deviation of the outcome variable which is one of the color metric provided given values for all other predictors .For each color, a threshold value for deviation were given. From the predicted deviation, it was used to identify the possible value of deviation which could occur in a particular situation. This data was used to adjust the values of significant predictors. Tools and Technologies: Python, R… Show more

Data from machines were given and we created a regression model to predict the range and deviation of the outcome variable which is one of the color metric provided given values for all other predictors .For each color, a threshold value for deviation were given. From the predicted deviation, it was used to identify the possible value of deviation which could occur in a particular situation. This data was used to adjust the values of significant predictors. Tools and Technologies: Python, R, Regression Models (Multivariate, Ridge, Lasso, KNN, Tree Models), Classification Models (Logistic Regression, Random Forest), Deep Neural Networks, Scikit-Learn, SQL Developer Show less

Analyze the data generated from the machine and predicting the time to the occurrence of next event for the machines. Data which occurs during the time of failure and normal conditions are taken and from that, survival model was created by also considering the problems of censoring, which will to predict the time remaining for the next occurrence of failure. These results were used to plan the maintenance activities of the respective machines. Along with this, respective survival metrics and… Show more

Analyze the data generated from the machine and predicting the time to the occurrence of next event for the machines. Data which occurs during the time of failure and normal conditions are taken and from that, survival model was created by also considering the problems of censoring, which will to predict the time remaining for the next occurrence of failure. These results were used to plan the maintenance activities of the respective machines. Along with this, respective survival metrics and performance of different models of machines were analyzed. Tools and Technologies: Python, R, Survival Analysis (Kaplen Meier, CoxPH Regression), Censoring, Scikit-Learn, Survival, SQL Developer Show less