Feb 2021 – Sep 2024 · 3 yrs 7 mos

• Enhancing text search relevance
• . Developed an NER model to extract the product names and key entities from metadata of learning
• objects (course, certification, etc.)
• Applied sentenceBERT model to vectorise learning objects and the searched query to retrieve
• candidate results
• Utilized entities to boost the individual scores of keywords in Elasticsearch queries.
• Used Reciprocal Rank Fusion scoring method to rank the combined results from Elasticsearch and
• vector search
• Improved overall user satisfaction score from 66% to 81.2% .
• Search Usage increased by 42% in two months
• AI assistant for course authors & learners
• Used an LLM to assist authors in course creation by sug gesting course titles and descriptions;
• automatically creating quiz modules for the course
• Implemented a RAG chatbot to assist learners in answering their questions about the course content.
• Deployed Whisper model to generate transcript for course videos to enable GenAI workflows
• Deployed a vector database to store embeddings for the content
• Designed most effective prompts using various prompt engineering techniques for the LLM tasks
• The share of course titles and descriptions generated from the AI assistant reached 23% of the newly
• created courses

Feb 2019 – Jan 2021 · 1 yr 11 mos

• Recommendation Engine
• Designed and productionized a course recommender system for ALM to replace the old rule-based
• system
• Analysed 100 million+ enrolments to define relevant feature set for recommendation algorithm
• Used various feature engineering techniques to select most important features from course
• metadata, enrolments and user profile
• Trained a Learning to Rank model to correctly rank the candidate course.
• Implemented ETL pipelines to transform raw application data into the model input features using
• Apache Spark.
• deployed a graph database (Neo4j) to efficiently generate candidates for ranking algorithm. The
• existing RDBMS was not able to support the queries needed for our use-case.
• The number of enrolments into courses through recommendation strip increased by 63% within a
• month.
• Skill Tagger for courses
• implemented a Skill tagger for courses by training a doc2vec model
• Leveraged Wikipedia and Wikidata to normalize and clean a purchased dataset of industry skills and
• job roles
• Designed an algorithm using wikipedia hyperlink network, external skill dataset and doc2vec
• embeddings to map skills correctly to learning objects
• achieved a recall of 76% on the test dataset

Feb 2017 – Jan 2019 · 1 yr 11 mos

• Classifier for content curation in discussion boards
• Trained a multiclass classifier model (FastText) to predict whether the content of a post by a user is
• strictly on the topic defined by the admin for the discussion board.
• used wikipedia and other open sources to create a labeled training set for the model
• achieved a precision score of 93% on test dataset and 84% on production.
• Was granted two US patents for the innovative training method which we designed for this model

Jun 2015 – Jan 2017 · 1 yr 7 mos