Apr 2022 – May 2024 · 2 yrs 1 mo

• I closely collaborate with the product team within the Discovery Search Science group to transform intricate business challenges into data science problem statements. My role revolves around enhancing the search experience for users by optimizing multilingual search outcomes across a wide range of customer products and internal tools and also engage in both offline and online assessments with the engineering team to ensure the feasibility of our solutions, utilizing methodologies like NLP and Rule-Based systems.
• Retrieval Augmented Generation (RAG):
• Involved in development of Retrieval Augmented Generation for search and chatbot applications. This entails retrieving academic content with ontologies information and dynamically selecting prompts to generate contextually relevant responses from a generative model.
• Multilingual Hybrid Search:
• Developed search capabilities in 11 Indic languages, including query understanding (QU) and query expansion (QE) modules, enabling users to seamlessly search in any language.
• Conducted benchmark evaluations on various vector databases, including Milvus, Qdrant, Elasticsearch, and Solr, with a primary focus on retrieval latency and vector index types.
• Integrated a fine-tuned embedding model into the inference server and incorporated a vector database into the hybrid search pipeline, enabling semantic search capabilities.
• Established a search feedback pipeline using the Gradio Interface for SME validation and the evaluation of various search algorithms.
• Setup search utilization dashboard by consuming user event logs to monitor and measure search performance metrics.
• Achieved a substantial 8% increase in Click-Through Rate (CTR) by optimizing the intent and entity based ranking algorithm, contributing significantly to the system's overall performance.

Dec 2021 – Apr 2022 · 4 mos

• Entity Extraction (NER):
• Designed and implemented a scalable text entity extraction algorithm integrated into a microservice. This system identifies academic entities through the utilization of multiple ontology datasets, synonym dictionaries, and a spellcheck mechanism, all seamlessly integrated with the Solr analyzer to achieve heightened accuracy.
• The algorithm serves as a core service in various client products, tasked with highlighting academic entities and retrieving associated academic content.
• Spell Check:
• Spell corrector is fundamental component of our NLU pipeline and search engine. Worked on building a corpus for spellchecking using user search query data. I implemented a hybrid spellcheck algorithm that combines popularity-based and context-based approaches. This algorithm successfully improved the accuracy of our system, resulting in the correction of approximately 37% of total user search terms with an impressive accuracy of around 79%.

Sep 2021 – Dec 2021 · 3 mos

Jun 2021 – Sep 2021 · 3 mos · Ahmedabad, Gujarat, India

• ‣ worked on building a real-time telephonic conversational-ai system for an effective interview pre-screening round.
• ‣ Improved 𝐬𝐩𝐞𝐞𝐜𝐡-𝐭𝐨-𝐭𝐞𝐱𝐭 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 𝐟𝐨𝐫 𝐈𝐧𝐝𝐢𝐚𝐧 𝐄𝐧𝐠𝐥𝐢𝐬𝐡 𝐚𝐜𝐜𝐞𝐧𝐭 with transfer learning on the deep-speech STT model.
• ‣ Microservices were developed to evaluate the Montreal Cognitive Assessment (MoCA). The microservices contain a computer vision-based cube and clock 𝐝𝐫𝐚𝐰𝐢𝐧𝐠 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐬𝐲𝐬𝐭𝐞𝐦, as well as a context-based 𝐚𝐧𝐬𝐰𝐞𝐫 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐬𝐲𝐬𝐭𝐞𝐦 leveraging NLP concepts.

Nov 2020 – Jun 2021 · 7 mos · Ahmedabad, Gujarat, India

Jun 2020 – Jul 2021 · 1 yr 1 mo

Aug 2019 – Jun 2020 · 10 mos

Jul 2018 – Aug 2019 · 1 yr 1 mo