• There are a lot of PDF documents being uploaded to the stock exchange that are called bonds. The complexity of these bonds varies document to document, the use case is to automate the process of required information retrieval using ML techniques as on date the work is being done by SMEs manually. (Note: there are more than 400 fields)

The information retrieval is of two types

• Entity extraction- fields such as date amount etc that can be directly extracted from the document by a custom NER model
• Sentence classification- fields that require class assignment based on text.

My Responsibilities and contribution:

• PDF processing: we need to split the PDF into sentences so that the sentence classifier can classify the sentences into desired categories. I was responsible for making the splitter, I used the Pymupdf library.
• Sagemaker pipelines for classification and extraction I was responsible for the classification pipeline and decided to use the setfit model which is a few-shot learning deep learning model, even though the model produces very good results the actual challenge came from a sheer number of fields in the scope(>400) hence I decided to get creative and combined several fields together and used capabilities of Setfit (differentiable head, head freezing) to combine multiple fields and use logistic regression head over each field to find final classification.
• Hyperparameter tuning We know that every ML solution requires hyperparameter tuning The dynamic nature of the use case entailed an organic approach for sentence classification I decided to use the optuna backend and hyperparameter search method provided by the setfit library.

• The project was to make an AI agent that uses a genetic algorithm to play with a human a game of killer sudoku in which a random grid of nXn is initiated and one killer number is selected, both human and AI will have 3 chances if they select two adjacent cells with total equal to the killer number.
• I was able to make that agent using end-to-end genetic algorithms implementation.

• To build a predictive model that answers the question: what sorts of people were more likely to survive? using passenger data
• I was able to use various ensemble and deep learning methods in this Challenge to achieve 1100 rank out of 14000 teams.
• This project involves the basics of feature engineering and various machine learning algorithms with hyperparameter tuning.

• Twitter has become an important communication channel in times of emergency.
• The ubiquitousness of smartphones enables people to announce an emergency they're observing in real-time. Because of this, more agencies are interested in programatically monitoring Twitter (i.e. disaster relief organizations and news agencies).
• But, its not always clear whether a persons words are actually announcing a disaster. the task is to build a machine learning model that predicts which Tweets are about real disasters and which ones arent.

• The TCS government project group made a product named GovFramework this can be leveraged to make any project and dependencies were added at run time.
• I was part of the framework team and was able to devise a method that uses property files to change the runtime dependencies.
• I was awarded 11000 TCS gems for this contribution.