Aug 2023 – Jul 2025 · 1 yr 11 mos

• Part of the foundation team of Splore for building the best human centric answer agent. I have designed and deployed a RAG-based, multi-stage, multi-agent support bot enriched with a knowledge graph.

Feb 2021 – Jul 2023 · 2 yrs 5 mos · Bengaluru, Karnataka, India

• NLP| Speech| Search| Information Retrieval| Recommendation| Research
• As part of Discovery search science team my task is to work with product team to understand business problem, transform those as data science problem statement doing research and come up with best possible approach, analyze the data, create data pipeline and develop state of art Models, design Deploy ML Model and Run ML and AB Tests. I also take care of offline and online evaluation along with Engg. team and make it deploy-able. The Approaches involves NLP, Rule Based, RL methods.
• Projects
• Spell Correction:
• Spell correction is one of the fundamental service in search engine that goes before any other service and it make it very critical .Worked on creating data pipeline doing feature engineering insights from this. Created Neural Network based spell correction model which improve performance by 75% from the existing implementation and evaluated offline and online via AB test.
• semantic Product Search,
• Retrieval using traditional methods utilised laxical token matching. That comes with multiple limitations
• To enhance the recall of relevant products, even when a token match is not present, I have implemented an advanced technique called extreme multi-class classification. This approach allows me to retrieve the most relevant products for any given in- stance, even when there may not be an exact match based on individual tokens.
• Query Expansion:
• Query expansion helps to add additional token in the user query that increase the recall of search, Create combination of models that includes Pseudo feedback based and Neural Network based and synonym based Approach to increase the recall of relevant search.
• Cnonicalization of query term:
• Query canonicalization is one of the method that is being used to improve user experience.
• Skill set: Python, PyTorch, GCP, Pyspark, Pandas, NLP, RL, Data Structure, Big Query, VS Studio.
• Keyword: Search, User Experience, Recommendation, Ranking

Jul 2018 – Jan 2021 · 2 yrs 6 mos · Bengaluru Area, India

• I worked as a Senior Data Scientist on the WWIFO team at NVIDIA. In this role, I worked on various machine learning projects spanning healthcare to fintech. My responsibilities included conducting ML research, optimizing ML models, and handling large-scale distributed model training and inference on NVIDIA hardware architectures. These proof-of-concept demos are utilized by the business and client solution teams to during business strategies. I also being part of business strategies. Ad-
• ditionally, I led solution architects and data science interns and published research
• papers in top conferences.
• Projects
• Indian Smart cities:
• Led the India Smart Cities Proof of Concept initiative, where I developed and implemented advanced computer vision algorithms for object detection, tracking, and segmentation to enhance urban infrastructure and services.
• Recommendation:
• Worked on enhancing click-through rates for the food industry. Analyzed industry data and trained an optimized XGBoost model, resulting in improved convergence and performance.
• Anomalies detection:
• COVID Detection with CT Scan: Developed a solution for
• COVID detection using CT scan imaging to assist radiologists with initial filtering.
• Text Summarisation:
• Developed solutions for legal document summarization in mixed languages, question answering, and machine translation using BERT-based sequence-to-sequence modeling techniques.
• Text to Speech:
• Indic Languages Automatic Speech Recognition using MetaLearning Approach. Conformer-based models have recently outperformed transformers in ASR, while meta-learning aids deep learning with limited data. We use Conformers to model audio dependencies and apply meta learning for rapid adaptation to new languages via the MAML algorithm. Evaluating on seven Indic
• languages, our approach, MAML-ASR, nearly matches state-of-the-art monolingual ASR performance in character error rate.
• Research Collaborations:
• Collaborated with Academia (IITs/NITs) for research work.

Aug 2017 – May 2018 · 9 mos · Bengaluru Area, India

• As part of research Internship built a Deep Learning model to do Identification Of Known Faces From Video.
• Identify the possibility of presence of known faces in video frames with highest sensitivity.
• Generate metadata about the location and time instance and possible identity about the known faces.