Joining Webex as a machine learning engineer: An interview with Ritvik Shrivastava
Here on the MindMeld team, part of the Webex Intelligence team at Cisco, we’re building technology to improve the traditional office experience in a post-pandemic world. One big priority will be touchless interfaces, which we’re able to power through Webex Assistant, theAlso enterprise voice assistant we launched in 2019.
Webex Assistant allows customers to use their voice to control Webex video conferencing devices, schedule meetings in a room, join online meetings, call people in their company directory, and more. It also provides in-meeting support for things like creating action items, taking notes, and scheduling follow-up meetings.
In 2019, we open-sourced the MindMeld conversational AI platform that powers Webex Assistant to allow anyone to build these types of conversational systems for other domains.
Behind these innovations is a team of engineers working to constantly improve our AI technologies and products. We caught up with Ritvik Shrivastava, a machine learning engineer who recently celebrated his first anniversary with the MindMeld team. Ritvik joined the team after graduating from Columbia University and interning with us in the summer of 2019.
Hi! Can you introduce yourself and share with us what your current job and responsibilities are?
My name is Ritvik Shrivastava, and I’m a machine learning engineer on the MindMeld team in the Webex Intelligence group. I work on designing, developing, and deploying intelligent solutions to the open-source MindMeld conversational AI platform and the wider Webex Intelligence services, like Webex Assistant. Our team focuses on applied machine learning and all its modalities: vision, speech, and text. My focus is primarily on text-based systems or, more broadly, Natural Language Processing.
What inspires or motivates you to build products at the intersection of ML and NLP?
Out of all the modalities within AI and machine learning, I’m most enthusiastic about natural language processing with text as the core input type. I believe that growing up reading a lot and writing my own prose and poetry is the main reason why I found it very natural to transition into linguistics, and later NLP, once I had forayed into the unknown world (from the eyes of a sophomore) of AI.
This field’s premise always made me wonder: what would it take for machines to achieve the same human-level understanding of writing and comprehending? With recent advents into natural language generation, I think we’re getting closer to that reality (but not quite there yet)! This gap that is yet to be bridged is one of my main motivators to work on the intersection of NLP and machine learning.
How did you find out about MindMeld, and what were your initial expectations when you chose to work here?
Conversations are the primary mode of communication for humans. Furthering our understanding of them and improving how machines perceive them is an important research goal in automating the field of computational linguistics and bringing us closer to the expected image of AI. In graduate school, my research was focused on understanding discourse, argumentation, and speech processing – all components of conversations.
At this point, I was also looking for opportunities in the industry to work on real-world applications and scalable means to implement similar research ideas. I learned about MindMeld through a friend who had previously interacted with the team. Once I applied and spoke to some team members, I loved the idea of working with them. MindMeld’s open-source focus on creating a single-stop NLP architecture that allows anyone to develop their own smart conversational apps excited me the most. The team was also developing Webex Assistant, one of the first industrial smart assistants, which was a very rewarding product to work on. Once the offer came, it was a no-brainer to accept!
What were some challenges you faced while developing for the MindMeld platform for the first time?
While having been involved in research for a while as a student, working on the MindMeld platform was my first foray into developing and deploying industry-level solutions using AI. The initial experience came with its own challenges. Not only is the focus on novelty, but also scalability, latency, and performance of the deployed ML models. Over time in my initial weeks, I learned how important these factors are towards providing end-users the ability to use state-of-the-art AI technology without costing them a fortune. This was definitely the biggest learning curve for me and is a skill I am still continuously trying to hone.
Are there any interesting, unexpected challenges that you’ve encountered?
Not unexpected, but one extremely interesting challenge is to work on internationalization or support for multiple languages through the MindMeld and Webex platforms. While massive amounts of research in NLP conferences is done on English language data, there is an increased awareness on developing tools and methods that can be expanded to multiple languages. This is extremely important to make advanced software systems accessible by people from all over the world. I have really enjoyed seeing our team’s focus leaning towards that. Inclusivity in terms of languages has been a key goal over the last year and factors into every new feature we choose to develop for the MindMeld platform, such as the data augmentation and annotation tools. This multi-lingual support is also adopted by Webex Assistant. My curiosity to speak to the multi-lingual assistant resulted in another great development for me personally: learning Spanish, or starting to.
What does a typical work week look like for you?
The work week balance on the team is pretty good. With only necessary meetings scheduled, everyone has ample time to work on their weekly targets without having to spend extra time outside the standard working hours. This allows for a great work-life balance and gives me enough time to enjoy all other activities without any sense of work-related stress.
What problems have you been focused on solving at MindMeld lately?
Recently I’ve been focused on adding automated data-augmentation capabilities to MindMeld that will allow developers to augment their training data multiple folds and make their conversational apps much more robust. Along with my colleagues, I’ve worked on integrating multi-lingual paraphrasing capabilities to the MindMeld platform using state-of-the-art systems.
I’m also working on active-learning-based query selection and auto-labeling with weak supervision, which will enable developers to automatically label queries into respective domain/intents and select the best-performing ones out of the lot. This will help keep the size of their applications to their chosen size and still perform efficiently.
Besides your main work, what other aspects of work life at MindMeld do you enjoy the most?
Throughout my internship with the team, and ever since joining as a full-time employee, I have enjoyed how sociable this team is!
Before the pandemic, when we were in the office, the happy hours, movie nights, small breaks to play games, and daily lunches under the bright San Francisco sunlight were nice breaks from work and helped to get to know everyone a little bit better.
In our work-from-home environments, we sometimes have virtual game hours that end the week on a fun high! We have had really interesting discussion spaces over the last year on sharing our cooking experiences, DIY creations, and my personal favorite, the push-up club! Yes, we did push-ups as a group in the middle of the workday on a video call, and it was excellent.
Working from home is often challenging. What has your remote work experience been like, and how did you feel about the transition from a physical office to WFH?
Several people on our team worked remotely even before the transition to WFH due to the pandemic. While this allowed us to have a good understanding of having regular meetings virtually, it also helped that those team members could provide great insights into making WFH productive. This created a relatively easier transition for me to WFH after the initial few weeks.
Our video is turned on for most of our meetings, making them feel more interactive and helping to avoid an isolated behind-the-screen experience. Working with Webex and other collaboration tools, it’s been easy to reach out to any teammate whenever I need to speak to them, so that personal connection has been maintained.
As for challenges against monotony, the WFH culture on our team provides good flexibility to schedule my work throughout the day and week so that I can get my work done in the most productive way while taking care of urgent or personal errands when the need arises.
Are there any interesting work traditions at MindMeld?
There are a few! Our weekly standups on Mondays are followed by optional weekend updates. They are a great medium to learn about your colleagues’ interests and lives outside of work. I’ve had some great conversations through these weekend fun updates and learned about many new movies to watch, places to visit, restaurants to explore, and much more.
We also have a fun bi-weekly tradition of the Big Hero award. Inspired by the movie Big Hero 6, this award goes to the biggest hero or the best performer from the past couple of weeks, decided by the award’s previous winner. It’s nice to see everyone get recognized by their colleagues and see how supportive the environment is. While right now it’s a verbal award, in the pre-WFH times, the award was an adorable Big Hero soft-toy. Definitely worth the hard work.
The domain of NLP is rapidly evolving, and a lot of academic work has been produced every quarter for the last few years. How do you manage to keep track of the latest happenings?
We have two reading groups in our teams: one for machine learning-focused papers and the other for user experience and metrics. These are great ways to keep up with the latest research in our team’s domains, and even ones outside that scope. Continuously learning about novel ideas and discussions through these sessions provides key insights into how we can develop these ideas into tangible industrial solutions and keep us updated on any relevant happenings.
What are some exciting research developments and industry use cases that you and the MindMeld team will be keeping an eye on over the next year?
In NLP, we’ll be looking out for the latest developments in the space of transformer models as state-of-the-art classification systems, data augmentation techniques for conversational domain data, natural language generation systems to artificially generate dialogue, and systems that will allow us to better expand to a multi-lingual setup. At Cisco, we have a strong focus on making our products accessible to everyone, and with that in mind, we are looking to extend the support of Webex Assistant to multiple languages around the globe.
We also want to make it easier for our users to identify action items, reminders, and key takeaways quickly and automatically across all the communication channels they use on a day-to-day basis (messaging spaces, emails, meetings, etc.). With this in mind, we’re looking at ways to employ NLP to help users gather these essential highlights.
Apart from NLP, our broader team works on speech processing and computer vision as well. We’re working to continually improve the user experience for features like meeting transcriptions, gesture recognition, and virtual background replacement. Across the team, we’re also looking into model and pipeline security using differential privacy and other privacy-enforcing techniques.
Interested in joining the MindMeld team? Send a mail to [email protected]!
Click here to learn more about the offerings from Webex and to sign up for a free account.
Sep 21, 2021 — Jamie Romanin
Sep 16, 2021 — Aruna Ravichandran