AI-Powered Customer Support with HelloVera - #18

Key Takeaways

[Music] hello and welcome to another episode of twiml talk the podcast r i interview interesting people doing interesting things in machine learning and artificial intelligence I'm your host Sam charington this time around we're doing something special here on the show and I'm really excited to share with you my producer and I spent some time in New York City this week guests of the fine folks from the future Labs at NYU Tandon and FF Venture Capital who are sponsors of this week's show and the organizations behind the NYU ffvc AI Nexus lab startup accelerator program you might recall that I mentioned the future Labs AI Summit last week an event hosted by the AI Nexus lab to showcase the startups in their first batch as well as the impressive AI talent in the New York City ecosystem we attended this event on Wednesday afternoon and had an amazing time we got to give out a bunch of twiml stickers meet some fans of the show and check out a great set of speakers I particularly enjoyed watching AI luminaries Gary Marcus and Yan laon both NYU faculty and longtime friends taking swipes at one another from the stage it was great I tweeted some of the highlights which you can check out by looking me up on Twitter at Sam charington we had our mobile Studio set up in a backstage dressing room and over the next few weeks we'll be sharing some great discussions we had with some of the speakers at the event but first we've got a monster of a show for you this week direct from the future Labs incubator space we spoke with Founders from the five companies in the inaugural batch of the AI Nexus lab program hello Vera clera behold. Cambrian intelligence and Alpha vertex these companies are doing some really interesting things and I think you'll enjoy hearing their stories we also spent some time chatting with Steve Kuan the managing director of future Labs about what makes the AI Nexus lab program so special we're releasing all six of these interviews simultaneously for your binge listening pleasure as twiml Talk number 18 Parts 1 through 5 before we get started though if you like what you hear you might be interested to know that applications are now open for the next AI Nexus lab cohort the program gives companies $100,000 in investment Capital four months of customized programming plus three additional months of free space along with curated mentorship from leading NYU academics and New York City investors pro bono services from Partners like lawyers and designers plus a paid for student fellow for the duration of the program additionally each company is partnered with an AI Nexus lab corporate pilot partner which in the last cohort included Daimler at futur labs. NC with any questions and now on to our [Music] show hello everyone this is Sam charington from this week in machine learning in Ai and I am live here at the NYU futur Labs accelerator space uh and I'll be talking to several of the companies from the NYU AI Nexus accelerator and the first up here is hello Vera and I'm here with uh the I'm here with James fan and leang leango uh guys why don't you to say hi Hi Sam hi ran so uh tell us a little bit about hello Vera yes uh hello Vera we are applying artificial intelligence technology to the problem of customer service um as you know customer service is a major pain point for a lot of Enterprises um it requires a lot of human and agent it's very costly operation often it costs $2 for a human agent to send you email and $5 to answer a phone call um with all the money spent customer satisfaction actually is alltime low um so last year I went to talk to L and said you know what this is a problem that artificial intelligence can help a lot um you know what makes human hate to do this job is this is very repetitive if you look at the top customer service issues they're all almost always the same issues repeating over and over but that's a perfect fit for for AI because the repetitive repetitive nature makes it easy for AI to learn to recognize the patterns in it and then there's also a lot of data in this domain every company keep track of the previously solved tickets and how human agents have solved the tickets you know data is the lifeline of artificial intelligence right so there's data we can use machine learning technology to learn how to resolve it just like human would um and what's better is when once you deploy you have additional new data coming in so the AI technology for customer service will get better and better after it gets deployed uh so you're you're part of this accelerator now how long have you been doing this uh we started last year so we just about a year okay uh and there have been a number of companies going after the the customer support um the customer support Market um and trying to apply machine learning and AI everything from startups to you know big companies in a space like zenes and others what is unique about your approach right um so if you look at the space in this um there are three type of companies there's one type that's doing um essentially a frequently Asked question search type of things um you know that's has a limited capabilities a second type is a what we call agent assist basically they provide tools to make human agents to do their job more efficiently that's nice but your human agent still the bottleneck um if your human agent go home your customers won't get their questions answered until they come back again um the third type is uh these chat platforms that if your company have enough technology techies available you can tell them to use the these chat platform Builders to build a chat bot but one thing is very important to know is most of these chat Bots are very script based right so if you have scripts then one thing you find out very quickly is once you deploy customers will never behave the way you imagin they will they don't follow the script and then your chatbot just falls apart yeah so our approach is we are taking machine learning data driven approach so it's a lot more robust um um and as we have more and more data it actually gets better and better and our bot actually is autonomous it gives answers in real time instantaneously okay can you um tell me a little bit more about that I mean I think most people who are doing Ai and machine learning would characterize it as data driven What specifically are you doing to collect and process this data to uh make your Bot smarter yes uh we have uh come up with a proprietary machine learning algorithm um based on deep learning it's able to beat a number of um deep learning based approach State art deep learning approaches um a number of natural language processing tasks as well as a number of Benchmark data sets um our deep learning program is able to take advantage of all all the data for in customer service domain um even starting with a small number of data but its performance continues to improve as the data gets get more and more available H how much does how much data does a customer typically need to have to start off with we can start off with very little data we can start off with just a dozen two dozen instances um with what's an instance a dozen a dozen or two dozen question answer pairs let's say okay and then we are able to start with running and and produce answers when you're so you start with your couple dozen question answer pairs as you uh increase the amount of data do you also have to increase are you also increasing labeled uh data or do you have a way to do kind of online labeling right we we are actually taking advantage of um users response by observing user response to infer the the label uh automatically and then we can able to take advantage of that okay so you mentioned uh a bunch of Technologies deep learning NLP uh and others how does how does that all fit together to allow you to do all this um so this is actually I guess it's fairly obvious to us because we've been working in this space for a long time um my background is in question answering um L's background is machine learning okay um we met at IBM several years ago okay uh as researchers where you used to be at IBM research um and New York or uh in Yorktown right here outside of New York um so when I was at IBM one of the things is I was uh one of the main leads for this IBM Watson question answering system that um you may have seen on TV that Bing at Jeopardy okay um so I've been working in this question answering domain for a long time so one of the things for us for in order to solve customer service issues you need to figure out what customer wants understand what customer want want and the delivered answer they were looking for so in order to do that you need to use a variety of natural language processing techniques as well as using machine learning techniques to recognize understand what customers needs are and figure out the right answer M and so uh let's take uh one of those deep learning for example how exactly are you um does deep learning play into what you do yes so deep learning is very nice in term of can take full advantage of all the available data to deliver high performance and as I mentioned you know in customer service you can actually get quite a bit data compared to a lot of other natural language processing tasks often you can have um supervised data based on the prior approach and we are taking full advantage using deep learning to uh deliver a better performance mhm and so did you did you base your deep learning models B on um you know some of the Contemporary industry research or did you develop them all from scratch what helped you get to where you are now um we are researchers so we are very familiar with the latest trends in the research Community latest uh Works um the specific um deep learning approach is what we have come up on our own um yeah there are lot lot of things we can see about Deering probably the time is not limited well let's let's get started yeah just self-promoting uh I have a class teaching I'm teaching a class at Columbia University on deep learning all the slides online so welcome to have a check okay Columbia class what's the name of the class uh the Deep learning for uh language speech and anation and Innovation vision and vision language speech and vision got it got it um yeah and I'm I'm probing a little bit into how you guys are using deep learning because I find that the the listeners want like stuff that they can go and apply what would you tell folks that want to apply deep learning to building their own you know speech models um and um maybe even applying it to customer service what have you learned that someone else could go uh use right so um I don't know about speech in particular but for natural language processing the key issue is sparity um the you tend to have a very high dimensional sparse uh space you are working in and you want to use deep learning um you want to reduce this Dimension high dimension often you have million Dimension highly SP sparse um features you want to reduce into a lower Dimension dense feature space that makes that can take advantage of um other aspect of deep learning easily and what are the features uh in the space that you guys are working in uh we use a variety of those features we actually tend to get extremely sparse high dimension features um typical it's a word phrase uh other linguistic features etc etc and any particular techniques that you're using for the dimensionality reduction um we are using our own proprietary U projection technique okay for Dimension reduction okay and so how about on the NLP side of things um anything interesting you can share about your approach uh on uh with NLP one thing I can mention is um in order to um a lot of time you have you you want to start you want to have a co-art a co-art is very important um what it means co-art is if you have limited amount of data right uh you still want to use deep learning approach um a co-art can give you high performance even with limited amount of data and so to achieve uh the AB ility to do a cold start are you using transfer learning or are there other techniques that that you're taking advantage of um yeah we that's one of um the core technology we have it's the proprietary uh machine learning technique that actually can do co-star very well okay okay tell me where you are uh as a company and where the product is as a company are you generally available um so our customers are large Enterprises um usually we talk to customer service directors in large Enterprises we currently have two pilot Partners we work with Okay uh and then we previously licensed our technology to a Fortune 500 company for their in-house uh intelligent agent development okay and what can you what can you say about the kind of the process for going into a new customer um when you you know they they say hey this sounds great we want to deploy this what does that look like what do they have to do actually we make this process really simple we we go give them a demo we by time we talk to them we already have a demo using their data okay their publicly available data up and running so they can interact with a agent bought for their domain by the time we walk in the door what does that mean they're publicly available data most companies don't have like their support all they so for example think about their tweets ah okay they they um a lot of companies provide customer service through Twitter right and that data is public and we can show how an agent behave on tweets based on their existing human agent tweets okay um so then we already have something by the time we walk in the door and of course uh that's just a start right we have if we have more data from them uh do deeper integration our AI technology can do a lot more okay and the AI will will become smarter and smarter once he learn to interact with customer and agents mhm uh how well does are there any issues associated with transferring uh or you know built training a model on social data and then sending it off to to go after emails does it transfer very well or there tricks that you need to play so there is a little bit um the language usage uh if you will right if you look at the the way people write tweets that English is slightly different from the way they write emails and you can even say um if people are on mobile platform then they tend to use short hands for words that you otherwise wouldn't say in um if you're typing out on a fulls siiz keyboard mhm um but other than that yes we will be able to handle this uh so you go in you are able to demonstrate some value to them up front based on the the social interactions and then they say let's do this and they want to uh pull in you know more of their proprietary data are they are you um able to hook up to their existing systems and what systems are those are those email or they something like a Zen desk or a support system or right so we do have this seamless integration with existing ticketing system they have whether that's zendesk Salesforce or um Oracle or any other ticking system so that makes it very easy for them to integrate with our system mhm and then do they have to uh do you have some kind of console that they use to label and tag different uh interactions or does it the system do it all for them so we provide a dashboard which gives them a um 30,000 ft overview of how things are going MH they can look at how many tickets are coming in how many tickets are resolved where the tickets are coming from etc etc and they can uh learn about what are the tickets that's not being resolved etc etc mhm and do you I I imagine that I'm thinking of like uh kind of you have all these tickets right you have some that are being resolved and um are you showing them is the system able to figure out uh and display confidence levels like you know we're pretty sure we nailed the these and not so sure how does that is that just displayed on the dashboard yeah so we they can sort by the the confidence they can sort by the time they can sort by a number of these criteria they can do the filtering to U do run report look at the tickets um there are a number of ways for them to um to to check to run analytics on on their um system and then I would imagine uh if you're doing this at scale another thing that would be important is um for the the things that you don't that you haven't been able to resolve correlating them so that you know we've got you know a thousand things here that we had no ability to help out with but we think that you know these 200 are all kind of about the same thing is is that something that you guys yeah that that's a great idea that's exactly the direction we are heading as well okay yeah maybe you should work with us contribute to the what other what are the other things that that you know on that topic what are the things that are you know either on your road map or more generally things that you think this space needs to evolve to be able to do in order to fully realize the vision for customers yeah so I think um I think we are in in the right space our claim is we think this space that we'll see more and more issues automatically resolved by AI imagine 10 years from now I'll say overwhelming majority of customer service requests will be automatically responded immediately only when you have rare cases that requires human attention and you'll take a little bit longer for human to come back and that that's that's across multiple channels it's not just email also chat social media texting etc etc mhm uh how did you of all the things that you could have applied your backgrounds to um you know deep learning NLP all that stuff is super hot how' you come up with customer service so here here's the story um how I thought always a story yes that's always story right um I bought this CPU from New Egg and the new egg has this price matching policy okay right so the price dropped so I called them I was on hold and then I talked to the agent uhhuh but that wasn't the first time I do this I've done this before this price price matching on new act before so I know exactly the set of questions they ask and exactly how this process works okay so as I was going through this I was was on hold and I realized you know as annoying as it is for me that agent on the other side he probably even more annoyed because every day he's going through the same process over and over and that's something a must be able to do MH so that's why I thought hey maybe hey I went to talk to hey maybe that's something we should use AI to to to resolve address this pain Point nice nice um tell me a little bit about um one of the things that companies that are um you know marketing AI Technologies to um lines of business like support you know have to Grapple with this issue of our we you know replacing people you know augmenting people you know what are the kinds of questions that your customers are asking and how do you help them think through you know what the role of AI within um you know with Automation in these kinds of environments yeah so our clients their number one concern is customer satisfaction MH and um and our goals are aligned our goal is to improve customer satisfaction if you look at the customer service if you can get a 24hour response time for email that's considered good and U only less than half the company can achieve that goal MH by having AI there we can reduce the response time significantly and improving the customer satisfaction um and from agent point of view what we're doing is actually making their jobs more interesting instead of doing the same thing over and over they get to do things different and more um more difficult more challenging customer service issues making it uh a meaningful job for those so can deliver customer satisfaction and customers okay so the reason why that the response times are so slow because they don't have enough and they can't afford enough customer support people so if you can help them use those people more effectively then it's a win-win for everybody yeah exactly right uh anything else you'd like to add or uh how can folks you know find you and learn more oh check out our website hello. a um and if you're available come to our uh demo day tomorrow Wednesday Wednesday okay great awesome well thanks James thanks young Lang um thank you s all right everyone that's our show for today once again thanks so much for listening and for your continued support a huge huge thanks to our sponsors this week future Labs at Yu Tandon and FF Venture Capital don't forget that the application deadline for the next batch of AI Nexus lab companies is coming up quickly visit nexus. to learn more or apply or email hello at futur labs. NC with any questions and of course don't forget to share your favorite quotes from this week's interviews to receive a twiml sticker we're starting to get pictures back from folks who are proud l ly displaying them on their laptops and they look amazing you can get yours by sharing your favorite quote via the show notes page via Twitter via our Facebook page or via a comment on YouTube or SoundCloud the notes for this show will be up on twiml a.com Nexus laab where you'll find a playlist with all of the individual interviews and links to all of the people companies and resources mentioned in the show once again thanks so much and catch you next time