Jul 2021 – Jun 2025 · 3 yrs 11 mos · Bangalore Urban, Karnataka, India

• Intent Model to identify customer needs on social platforms
• Trained a classification model that will tag queries asked by customers on multiple social platforms like X (formerly Twitter), Facebook and Reddit etc to Fidelity's internal taxonomy encompassing 108 distinct intents.
• Leveraged training dataset prepared by both human annotators and synthetically generated data via paraphrase generation using T5. Transformers based Few-shot learning approach like SetFit was used to train the classification model. This approach enhanced model performance significantly, achieving an impressive F1-score of 74%, compared to the existing model's 22%.
• In-house Machine Translation model for non-english social posts
• Led the development of translation models to address non-English social media queries from customers. These models can translate text between English and three major non-English languages (Spanish, German, French).
• Explored various transformers based pre-trained models from HuggingFace and fine-tuned these models on domain dataset leveraging AWS SageMaker. Each translation models achieved BLEU score ranging between 50 and 60, indicative of high-quality translation.
• Developed UI for translation capability utilising Streamlit and backend endpoints developed with FastAPI. The translation models were deployed using AWS SageMaker.
• NLP Bootcamp sessions
• Conducted approximately 20 one-hour sessions for team members, covering fundamentals NLP concepts like text pre-processing, NER, vectorisation till advanced topics like encoder-decoder, attention, transformer architecture etc. This training initiative helped team to elevate understanding of crucial NLP concepts.

Feb 2021 – Jun 2021 · 4 mos

Aug 2018 – Feb 2021 · 2 yrs 6 mos

• Embibe is an Edtech startup that helps students to improve scores by analysing student behavior and recommending personalized content. My role includes generating personalized test papers, assignments using a configuration like difficulty level, chapter distribution, question type distribution, question discrimination, concept mastery, etc.
• Design scalable architecture: Asynchronous request execution using Redis (message broker) and Celery. Improved throughput of APIs by identifying bottlenecks using time profiling and performing load testing. Reduced network calls and I/O operations using Redis cache.
• Item Response theory: I have implemented Neural Network based IRT using Keras in Python.
• Discrimination, difficulty level, and guessability of questions are calculated using attempts data by learning weights in NN models.
• Automatic test paper generation: I have worked on genetic and greedy algorithms to generate papers and assignments. Diagnostic tests are generated to check the ability of a first time user coming to the platform.
• Test quality score: We have to measure how good the generated paper is when compared with the previous year papers. We calculate a score using multiple parameters.
• Concepts - Machine Learning, NLP, Neural Networks, Text Similarity, Statistics, etc.
• Techniques - Greedy and Genetic Algorithm, Item response theory, etc.
• Technologies/packages - Scikit-learn, Docker, Keras, Flask REST API, Redis, kafka, Celery, locust, Git, Elastic Search, Mongo DB, Amazon AWS services like EC2, S3.
• Challenges - Build scalable APIs with high throughput and efficient algorithms to package content with high-quality score.

May 2017 – Jul 2018 · 1 yr 2 mos · Chennai Area, India

• Indix has 2 billion products crawled from more than 2000 e-commerce stores. My role includes
• classifying products across 6000 categories by using textual information.
• Training Data Preparation: Trained Word2Vec model on breadcrumbs present in Indix Corpus to generate mapping of indix category path to e-commerce store’s breadcrumb.
• Classification Model: Built classification model to integrate products coming from various taxonomies to Indix taxonomy. We solved this as a classification problem using SVM and FastText. Our classification system classifies products with a precision of 85%.
• Concepts - Machine Learning, NLP, Text Similarity, Feature engineering, Word Embeddings.
• Techniques - Word2Vec, Naive-Bayes, SVM, FastText.
• Technologies/packages - Apache Spark, Scikit-learn, Spacy, Docker, REST Api, Git, Amazon AWS services like EC2, S3.
• Challenges - Training data collection, Multi-Class Problems, Scalability.

Jan 2017 – Apr 2017 · 3 mos · Chennai Area, India

• De-duplication Of Taxonomy
• Indix has 6000 categories to classify products. Identified more than 500 duplicates pairs using
• TF-IDF, Word2Vec, Cosine Similarity and LSH nearest neighbors.

Sep 2016 – Oct 2016 · 1 mo · Jaipur Area, India

• Working on Text summarization and Topic modeling using Latent Dirichlet Allocation and Natural Language Processing techniques.

May 2016 – Jul 2016 · 2 mos · Gurgaon, India

• 1. Image Classification
• Developed classification model using SVM and Image Processing Techniques. Processed images to obtain feature space by removing unwanted regions from the images with the help of Support Vector Machines.
• 2. Text Processing
• Generated Word Cloud of trending keywords on web using Natural Language Processing and n-grams model.

Feb 2016 – Mar 2016 · 1 mo · Gurgaon, India

• Built model for separating tags(author name, title, year of publication) from input reference string
• using Feature Engineering, Machine Learning and Natural Language Processing.

Jan 2015 – Apr 2015 · 3 mos · Jaipur

• Conducted lab sessions of students and helped them in understanding the core concepts, debugging programs, and improving their code quality