Partner Director of AI for Microsoft Copilot Studio, this includes managing a large team of applied scientists working closely with engineering teams and product teams in the delivery of Generative AI capabilities as part of core Microsoft Copilot Studio. Some areas that my team drives are: advanced RAG infrastructure, reasoning and orchestration via agentic patterns and frameworks, natural language to code (SQL, PowerFX), SLM and LLM finetuning, e2e GenAI quality and compliance (Responsible AI).

- AI projects coordination for multiple Microsoft Dynamics product groups: Process Insights, Customer Insights, Customer Service Insights, Engagement Insights, Power Virtual Agents- Lead AI development of AI-based chatbot for Power Virtual Agents with GPT3 Retrieval Augmented Generation (RAG) prompt engineering.- Lead development of AI-based NL to code authoring experience of chatbots with GPT4.- Developed GPT3 OpenAI Task Mining model to summarize meeting transcriptions into event logs. Leveraged Dual Transformer Encoders for clustering of multiple recordings.- Implementation of LLMs integration into business applications (chatGPT, codex-cushman, davinci, curie).- Explainable AI area efforts and oversaw the implementation of Local and Global AI models’ explainability with Shapley and DeepLift for NLP models.

- Coordinate v-team of Responsible AI champions across Business Applications & Platforms Group- Drive guidance and education of Microsoft's Responsible AI principles at group level- Drive compliance implementation of RAI principles across Business Applications & Platforms Group- Cross-team coordination at Cloud+AI Division level of RAI efforts

- Drove Privacy compliance and guidance for Applied AI team- GDPR tracking and evaluation- Collaborate with Legal on Privacy consultations

- Timeseries generation and analysis from search logs for financial quarterly revenue correlation.- Automatic classification of search categories and product lines from search logs for timeseries decomposition.- Large scale pipeline in distributed system for web logs analysis and feature extraction for financial models.

- Convolutional Neural Network for ad detection on website screenshots.- Bi-directional LSTM recurrent network for sentiment analysis for offensive classification of text passages.- Website affinity on query model using Glove word embeddings and cosine similarity aggregated across website’s documents.

- Implementation of Boosted Regression Decision Trees models for Content Quality detection. - Design and development of universal adult document classifier. Cross-language n-gram language model design and multi-stream aggregation of adult signals.

- Lead overall adult filtering and experience strategy for Bing En-US and drove collaboration with global teams on adult leakage reduction.- Implementation of adult filtering data mining models for Adult Leakage under En-US market.- Creation of crowdsourcing application for adult leakage measurement.

- Implementation of Spam/Junk judgment platform integrated with crowd-sourcing using IE 9 C# toolbar.- Statistical analysis of variance on Spam/Junk judgments on scraped results across crowd-sourcing judges. Bias and variance decomposition of judgment error.

- Evaluation of multiple document conversion third party libraries as well as developed file type detection system based on Magic Words.- Implementation of PDF title detection machine learning algorithm using boosted regression decision trees and random forest.- Implementation of end-to-end document conversion pipeline verification system with Win/Loss analysis.

- Designed and developed automatic testing in production infrastructure and deployment pipeline for the Wrapstar project. - Encoding detection mechanism testing for Bing. Developed side by side judging tool to compare previous and new detection mechanisms.- Site specific and forum packages feature extraction P/R measurements, regression, and performance testing.- OpenGraph feature extraction P/R measurements, regression, and performance testing.- Developed BI dashboard for feature coverage tracking over web index as well as index size tracking.- Structured data extraction from 1D and 2D tables testing as well as development of prototype Fact Answer for Bing with synonym analysis and n-gram matching over query string.