Pamela Bhattacharya

Applied Research and Engineering Leader

Microsoft (July 2013 - May 2021)

AI capabilities powering Office

I was a Principal Scientist and Manager of one of the central ML teams in the Office organizations. Our goal is to leverage trillions of user interactions in our product to build intelligent service-based features across Office apps.

Following are the areas I drove on the team:

  1. Designer Intelligence for PowerPoint: We used several ML approaches that can propose the most suitable layouts for the content, intelligently crop images, as well as automatically recommend relevant icons and pictures. We built customized recommendation models on image and text embeddings as well as features from user history to personalize the recommendation.

  2. Intelligent assistance on MS Office: We powered the intelligence for the “Tell me what you want to do” productivity feature for the entire MS Office. When the user enters words and phrases in natural language about what they want to do next, our AI powered solutions enable our users to save time by eliminating the need to click through tabs and multiple actions to achieve the desired outcome. We used a language model and action graph built on user action history for these recommendations.

  3. Authoring assistance for Verticals: I started/led an incubation project to improve authoring experiences for different verticals. We built a custom generative model for a couple different verticals, iterated with internal MSFT customers.


Scheduler - a conversational scheduling assistant

Product scope and impact: Scheduler (aka Calendar.help) is a conversational scheduling assistant. I led the overall scheduling intelligence, conversational understanding models and infrastructure, and the core meeting scheduling workflow in Scheduler. Over three years, increased scheduling task automation from a predominantly humans-in-the-loop system to >95% using a diverse collection of ML techniques in natural language understanding (NLU), entity extraction, relevance understanding, and automated reasoning. For more information, watch my presentation at Microsoft Faculty Summit or tune in to MSR podcast. I grew the team from 3 ICs to 25 ICs (ML engineers, Applied ML scientists and full stack engineers) as well as set the technical direction of the team for private preview and pre-GA.


I led the team find the balance between product virality and paid tier adoption. I was responsible for the health of our self-serve business, that required

me to flex between engineering, product and business strategy. I worked very closely with several business partners from Microsoft (Product, Research,

Design, Data, Finance, Privacy/Compliance) as well as enterprise customers (solution architects, enterprise productivity associates) to monetize the

advanced offerings.

Following are the areas I drove on the team:

Learning & Training: Models - platform / compliance; Data and telemetry pipelines; Data labeling and training pipelines

Reasoning : NLU - intent and entity extraction; Scheduling intelligence and personalization; Actions/decisions workflow; Human augmentation to system intelligence

Metrics and OKRs: User behavior data collection and reporting ; Product metrics; User feedback ; User Churn

Publications:

  1. ScopeIt: Scoping Task Relevant Sentences in Documents

  2. To Schedule or not to Schedule: Extracting Task Specific Temporal Entities and Associated Negation Constraints

Patents:

Granted: Signal analysis in a conversational scheduling assistant computing system

Filed: 13 patent applications

Intelligent Office 365 customer support

As part of an effort to improve Office 365 product offerings and customer service, we wanted to harness the various signals from free-form-text and predict anomalies in support tickets that could indicate product regression or system wide issues. Faster detection of these anomalies makes it easier for product groups to prioritize development work, as well as better customer transparency acknowledging that Microsoft has identified a service issue and working on fixing it. We used a combination of bag of words-based classification models as well as timeseries forecasting to detect anomalies in the service from support tickets.

Patents:

· Predicting service issues by detecting anomalies in event signal

· Correlating distinct events using linguistic analysis

Contextual self-help document recommendation

Recommending contextual knowledge base articles lets user’s self-help when they’re experiencing issues with O365 services (e.g. setting up domain, adding users, paying bills etc.) However, recommending all articles are not relevant to the user (similar to search engine results). Being able to mine through user activity logs and recommending them articles that could be related to the issue they’re experiencing is key to delighting customers. This also has a huge impact on support ticket volume and service Cogs.

Patent:

Context-Aware Display Of Objects In Mixed Environments