As a Principal Applied Scientist at Microsoft’s Maps and Geospatial Team, I have played a pivotal role in driving Location Understanding service to become a market leader in its segment. Key contributions include:Location Parser: Leading the enhancement, I successfully expanded its reach to over 15 languages, achieving significant improvements in token classification and reducing inference latency.To integrate heavy transformer models into a low-latency service, I developed a dual-model approach supplemented by a smart cache. This approach involved the use of a fine-tuned BERT model for tail queries and an ensemble of heavy weight models (fine-tuned Turing URL v6, DeBERTa, XLM-RoBERTa) for head and body queries. The implementation of mixed/half precision training further accelerated inference latency.To refine our existing training data, I utilized GPT Prompt engineering with Grounding and Chain of Thought technique for a location entity extraction task.To enhance the Parser’s domain knowledge, I added geospatial chain features and extracted high-impression knowledge graph entities, which encompassed geospatial locations across various categories such as Sports Teams, Person Names, Movies, and more. This signal notably increased precision by reducing false positive triggers by 14%.Relevance Re-Architecture and Stack Unification: I’ve built intricate data pipelines to augment training data and mine queries with high learning potential. Introduced contextual query understanding features, experimented with innovative training algorithms, and developed entity selection algorithms. I simplified and unified different region/market-based stacks, which involved phasing out outdated integrations and modifying the execution of key features.Technical Leadership: I successfully managed various projects as a Technical Leader, providing a clear, phase-by-phase roadmap. I mentored team members in Data Mining, Deep Learning techniques, and overall experiment execution.

As a Senior Applied Scientist with Microsoft’s Maps and Geospatial Team, I’ve been instrumental in enhancing Location Understanding across various search answers, including direction, weather, places, local, time zone, flight booking, and more. Key contributions include:Geospatial Cache Development: Leveraged Instrumentation logs and User Engagement metrics to construct a Geospatial Cache. This strategic initiative bolstered infrastructure performance, providing a latency buffer for deep learning experiments. The outcome was a significant latency improvement (~6ms at the 95th percentile) in both offline and online A/B testing, without compromising end-to-end retrieval effectiveness.Context-based Location Parser: Trained a deep learning sequence model to improve the overall relevance of the location understanding stack. The model identifies location and non-location tokens in a search query, resulting in a ~7.1% increase in tagging accuracy. This achievement was recognized in the Microsoft Newsletter and published as a research paper in MSJAR Vol14 (Microsoft Journal of Applied Research).Multi-Lingual Query Parser: Designed, optimized, and deployed a sequence tagging model by fine-tuning a 12-layer BERT uncased model. The model training engine was built using PyTorch’s Transformer library, and the acyclic computation graph was optimized using ONNX’s C++ libraries.

Have been a key developer in Ranking area since I joined the Bing Local Search team right after my undergrad in 2013. Under the able guidance of my managers and collaborative peers, I have been able to drive great results - working on different ML techniques. For instance, the ranker improvements have had a significant impact both on core Relevance metrics as well as end to end metrics like SBS (Side by Side). In addition to pioneering new techniques, I have been scaling these to multiple markets with limited effort. Designed and implemented a multi-class Taxonomy Classifier. Trained a stacked-LSTM model to deduce the taxonomy of the search query and employ that for Semantic Information Retrieval. Offline evaluation showed an accuracy improvement of ~4.5% compared with the production classifier. The work was widely accepted and I published a research paper in MSJAR Vol 12 (Microsoft Journal of Applied Research) and presented this at MLADS Fall 2019 Conference (Machine Learning, Analytics and Data Science).Developed an intricate Conflation model for “Deeplinks item – Local Entity Integration” and built an infrastructure to leverage Deeplinks ranking for improving Local Search Quality. The project required collaboration with multiple partner teams and was even appreciated by the Corporate Vice President of Bing Engineering.Have demonstrated leadership qualities in driving Relevance gap investigations (experiment design) to identify core work items and have been able to expand my influence on other team members.Have been an active member of Machine Learning Community, Organized Synapse’18 (Annual AI Meet at Microsoft India) and part of the core committee for AI School’19 – where I took a session on Sequence Modeling.Over the last few years, I have also been a part of Recruitment events, visiting multiple universities for hiring Software Engineers & Data Scientists.

As an Applied Scientist 2 at Bing Local Search, I have been involved in a variety of impactful projects:Machine Learning Model Exploration: I have explored various machine learning models such as gradient boosted decision trees, FastRank, and MART. My work has focused on mining high information gain features, such as user engagement and query/entity vector cosine similarity, and generating targeted training data to enhance ranker performance.Relevance Quality Improvement: I have added consequential impact to the relevance quality by successfully shipping numerous machine learning experiments for local search in markets including Australia, UK, and India.URL-Entity Extraction Design and Implementation: I designed and implemented a post-web Information Retrieval stack called URL-Entity Extraction. This involved creating a reverse index using local entity - primary website attribution and developing an infrastructure to employ web results for a search query to generate local entity candidates. This experiment notably improved local answer coverage by approximately 6.4% and was deployed following latency optimizations.High Inertia Initiative: As a key developer for the High Inertia Initiative, I ensured optimal relevance quality for epic local entities such as airports, universities, attractions, and others.Production DRI Team Membership: As an active member of the production DRI (Directly Responsible Individual) team, I have successfully developed tools to preemptively prevent potential live site issues during deployment.

As a key member of the Bing Local Search team, I thrived in an Agile environment, demonstrating my ability to quickly adapt and take on challenging markets such as Korea and India. My primary focus was on enhancing relevance quality and bridging gaps with Local Search Leaders.My accomplishments include the implementation of Dynamic Relax Radius Search (Information Retrieval), and the training of L1/L2 Rankers using 'learning to rank' methods. These initiatives significantly contributed to Bing’s successful launch in Korea.I was also responsible for deriving valuable insights from data analysis and identifying innovative techniques to enhance the search quality of Bing Local.In my first six months stint at Microsoft,I secured "First position" in the Machine Learning Competition held at Microsoft IDC, Hyderabad.