Marjan Josenia:
I had meetings with my manager, with my mentors, and I had the opportunity to get lots of information from them, and learn a lot from them because they're experience, initial learning and [inaudible 00:00:11] They gave me lots of insight when I look at the problem and look at the project, and help me, how can I approach a problem that is kind of challenging, and was very helpful for me, actually.
 
Hanslei Cuoscia:
My team is here for me and has my back, unconditionally. They just want me to succeed, so that feeling of the team just generally wanting your success instead of wanting their product built alone is really huge, right there. I think this is growing my career, significantly and there is no better internship that I've seen.
 
Barbara Alberts:

Curious about what it's like to be an intern at Salesforce? Welcome to the Salesforce Spotlight: A podcast series that highlights the extraordinary stories of the Salesforce employees and Futureforce interns. I'm Barbara Alberts, and on this episode of the Salesforce Spotlight, APM Intern Hanslei Cuoscia and PhD Intern Marchian Josenia talk all things Einstein and what it's like to work with artificial intelligence. Tune in for more, and happy listening.

So on this week's podcast, we have APM Intern Hanslei Cuoscia and Software Engineering Intern Marjan Josenia, and they're here to talk about their projects this summer and what they've been working on in their internships. Thank you guys for joining.

 
Hanslei Cuoscia:
Of course.
 
Marjan Josenia:
Thank you.
 
Barbara Alberts:
So first thing's first, can you guys just tell me a little bit about yourselves: where you went to school, wheat you majored in, things like that?
 
Hanslei Cuoscia:
Yeah, sure. So Hanslei, as you mentioned. I go to the University of California, Berkeley, right now, and I am majoring in Data Science with a minor in Computer Science, and I'm currently in the Associate Product Membership Program over here at Salesforce, and I'm on the Einstein Modeling Axiom.
 
Marjan Josenia:
My name is Marjan, I am majoring in Culture Science. I am a PhD student in the University of Houston. Currently I am working on improving semantic parsing for [inaudible 00:02:00] labeling in Salesforce.
 
Barbara Alberts:
How did you guys find out about Salesforce and about this internship program?
 
Hanslei Cuoscia:
I found out about Salesforce from a friend that was a Software Engineering Intern here, a year before I interned here in the summer. He told me great things about it, so I looked into it and then I got an interview. That's when I fully understood what [inaudible 00:02:21] business meant, and what all the clouds do.
 
Marjan Josenia:
Last year I attended Grace Hopper Celebration in Houston and was able to attend the career fair there, and Salesforce was there and I talked to the Salesforce people there and find out what Salesforce was doing, exactly. I liked the culture there because I have heard from my friends that Salesforce has an open culture, and I really like that.
 
Barbara Alberts:
Getting these internships is a pretty big deal. They really help give you a good foundation for what's to come. Why did you guys decide you wanted to intern at Salesforce?
 
Hanslei Cuoscia:

One main reason for why I picked Salesforce over other APM programs, and just other internships in general, was the things I heard about the APM program from the Inaugural Batch which was last year's APM program. What they told me was there's a ton of executive exposure and interaction with executive leaders, both in product and engineering, just all around. There are regular APM Thursdays, in which you and all the other 11 interns in your batch meet up on a regular basis, have workshops, meetings, and just question and answer sessions with a lot of these general managers, executive vice presidents and all these people that are literally setting and creating the vision for the next few quarters, or even years of Salesforce. I heard about that, I thought it was pretty crazy to be in a room with them, not only once or twice in an internship, but once a week. That was definitely a huge factor.

In addition to that, I also think that the network for our program is pretty awesome. By network, I mean that it's a 12 person program. Other programs are either one or two people for APM's, or maybe even a 50-person program for a couple other companies. I thought 12 was a perfect size to genuinely and personally know everyone else in the program, and just have a good time and be able build a quality, lasting friendship that would go years onwards, after this.

 
Marjan Josenia:
For me, I was looking for a challenging internship for the summer, and my majors actually: Mission Learning, DIP Learning, and Natural Language Processing. I was able to find out there are less options using [inaudible 00:04:20] Stat, or related to Surge, which I realize on the DIP Learning and Natural Language Processing. That's why I found that opportunity and I chose that opportunity at Salesforce.
 
Barbara Alberts:
Can you tell me a little bit about the projects that you guys are working on? You're both working within Einstein a little bit, but can you just expand on what, specifically, you guys are working on, and maybe how that relates to Einstein overall?
 
Hanslei Cuoscia:

This summer, I'm working on Einstein's first Business Card Scanner, meaning that you upload a picture of a business card currently as for web, and later on we're going to expand to mobile. After uploading that picture, then essentially all that information's parsed and pulled out of it, and you automatically have a new lead, record or a contact created in your Salesforce dashboard. Couple of the technologies use our Optical Character Recognition to pull the text from the image. Then, after the text is pulled, then we use Name Entity Recognition to pull the names, and locations, and organizations from that text. Finally, the end product is a new lead that's automatically added to your Salesforce records. So, that's what I've been working on this summer.

In addition that, one other thing I did was a Sentiment Analysis Lightning Web Component, meaning that it's just one of those components that you can easily drag around in your Salesforce set-up and embed it onto the screen, and you can literally have that tell you the sentiment, whether it be positive, neutral, or negative sentiment, of any customer's survey response, or any field that you want to get the general vibe of.

 
Barbara Alberts:
So, instead of having to go into Salesforce type in from this business card, the name and everything like that, it kind of automates that process?
 
Hanslei Cuoscia:
Exactly. Instead of creating a lead all manually by yourself, and filling out the five, to even 12 fields, you can just scan a card, and most of them will be filled out for you. You even have the option, as the user to edit whichever fields were incorrectly scanned, or even just completely add some fields from scratch. But it does most of the work for you.
 
Barbara Alberts:
Yeah. And Marjan, can you explain a little bit about what your internship project has been?
 
Marjan Josenia:
My project is about the instant search. In the instant search, we are giving two types of the queries: conceptual queries and non-conceptual queries. My project was to improve the parsing of the conceptual queries. The conceptual queries include some name entities, and in order to find out the name entities we need, and digital network that can identify the long names of the organization, the name of the persons, to create the final sequel. In order to retrieve the information that the user requests. So, my project was to improve that part, that name entity recognition.
 
Barbara Alberts:
Correct me if I'm wrong: so your project is a little bit more internally-facing, and you're working on an external feature, correct?
 
Hanslei Cuoscia:
Correct, yeah.
 
Barbara Alberts:
Yeah, so that's pretty cool. Can you guys tell me in your own words though, what is Einstein? I know people are like okay, it's AI, but can you expand a little bit on how that relates to your project, particularly?
 
Hanslei Cuoscia:

True. Einstein, in my opinion, is just the brain behind all of Salesforce's products. It's how the machine learning, embedded in every single cloud, or in every single functionality of the product that can potentially have predictions made, or can be a little smarter than just a simple row-based kind of system. That's what Einstein is, from what I've seen, but my team particularly creates 8-pix AI services, meaning that we create simple API's from which, with one line of code, you can easily have a machine learning API call, and on the Salesforce platform. Literally, in any field or any standard custom object, you can instantly get a response for translation, you can use that API. You can also use Sentiment Analysis. You can also use name Entity Recognition. You can also use Optical Character Recognition. So, these are all things that, with one line, you can easily utilize that power of machine learning. It can essentially be democratized for any Salesforce admin, any Salesforce user, or even an internal Salesforce user.

So, that's what I'm working on with Einstein, but as an overview, Einstein is essentially the brain behind all of Salesforce's products, and is embedding machine learning in a seamless manner throughout the entire platform.

 
Marjan Josenia:
I like to answer this question related to instant search, because from instant search previously, you were able to only search for the key words, but now you can search for the queries, which are conversational queries. So I just love having opportunities, you can say my opportunity last month, which is very similar to the conversational query. And that's what, actually, Einstein is. Convert query, maybe statical things, some very dynamic queries, and make the life easier for the in-users.
 
Barbara Alberts:
These queries don't have to be perfect, pristine, very formal rigid sentences, you can have a conversational sentence, go in there and kind of understand it. It's understanding the way that a person would actually talk, versus a textbook.
 
Hanslei Cuoscia:
Sure, casual and colloquial, instead of formal, right?
 
Marjan Josenia:
Exactly.
 
Barbara Alberts:
What were some challenges that came up during your projects you had to work through this summer?
 
Marjan Josenia:
The challenge that we had in the Lexus Word Search was that the name of the inquisition that are too long for us. 13 words, or maybe even 15 words. The current models cannot identify the name of the long [inaudible 00:09:37] So we needed to have a model that can identify them in order to search, convert the query to a sequel query. I tried to read lots of papers, and try different models, and I found, finally, model that can identify these long name entities.
 
Hanslei Cuoscia:

So, when modeling this business card scanner out, the name entity recognition that I mentioned earlier was definitely difficult to, not only get the names, not only get the organizations, but it was difficult to get the location and address from a business card. There's no way to easily identify the address, unless you have some very, very large database of addresses and some maps API that it's linked to, which we didn't want to go into because that would be outside of the scale of my internship. We only had around eight weeks to develop this product. So, keeping that in mind, we decided not to go with address right now, but we instead decided to optimize for the other fields that we could, in time, have a very high accuracy for. So that includes something that regular expressions can do, which is email addresses and phone numbers. In addition to that, we could focus more on fine tuning the name that pulled with the NER, or Name Entity Recognition API, and also the organization name, which was also successfully pulled at a high accuracy, as well.

So we optimize for those, and we had five to six fields that were with very high quality populated every time, any business card was pretty much scanned. We focused on getting a really good NVP out, instead of focusing on a very difficult and longer term problem, which was address, and I think that was definitely the right decision to do.

 
Barbara Alberts:
What are the next steps going to be for your projects once you guys hand them off, or once are at a place where you can hand it off, maybe not necessarily completed, but in a better place than where you found it?
 
Hanslei Cuoscia:

I think some potential avenues for my business card scanner product would be: there's a few. The first one could be that we could put it on the app exchange, which is Salesforce's apps. A second thing could be, we could create an unmanaged package for it, meaning that it's an unmanaged snippet of code that people can instantly run on their Salesforce platform without needing it to be actively maintained, or bugs-patched, or any of that. It's kind of like the [inaudible 00:11:45] where they can make modifications to it themselves, or if they feel like it's kind of lacking any features that they really need for their specific customer use case. And the third thing would be, we could create a managed package, which, based off of customer requests, we could add more features to it.

So, unmanaged package, managed package, or just put it on the app exchange as a whole. Those are three possible avenues for my project, and what that means is we, of course, want customers to use this scanner as much as possible. Zooming out even more, I think a next step, on the larger scale of things, would be to extend this to mobile, because business cards are generally scanned on phones. Because a lot of the technology, in terms of Optical Character Recognition and Name Entity Recognition, has already been developed over the course's internship, it's simply a matter of extending this use case to mobile and focusing more on the UI. Where as the back end has pretty much stayed the same. Because we've also done a lot of quality and metric testing and it seems like the quality's doing pretty well, so far. So it can easily be exposed to the public.

 
Marjan Josenia:
One part of my project was about how can we have more conversational queries and more compositional queries. So I think the second part of my project, I focus on parsing the complex queries. This is only the beginning of the instant search, and there are lots of rooms for improvement of this parse because when we see the user feedback, we see that they ask us to have more conversational queries and more complex queries. So in the second part, I try to represent the queries in a very different way, which is, I don't want to go into the details here, but just something similar that Amazon's Alexa is doing. Then, with this, we can add more flexibility to our queries, and help users with every query that they want in the instant platform.
 
Barbara Alberts:
One thing I'm genuinely just curious about is how do you guys handle foreign languages in these projects? Because Einstein and Salesforce have global products, and so how have you guys been working around that? Is there anything that you've been doing to expand on the abilities in terms of recognizing words or recognizing search queries?
 
Hanslei Cuoscia:

International relations is definitely a very, very important and high priority task for every single product on Einstein and throughout the company, as a whole. We really value our customers that speak different languages and that are in different countries. So, for my team particularly, this business card scanner will most definitely, with full intent, be extended to non-English speakers and people that have business cards in other languages. One quick way on how to do that is, remember how I mentioned that we have easy, simple, one-line machine learning services, right? What we're going to do is, first of all, we want to build the English product first, right? Simply in one language, to have an NVP, and approval concept that this business card scanner can in fact be made on the Salesforce platform, for [inaudible 00:14:33] And then after that, then we'll extend it to the next most popular language, which right now it seems like it will be Chinese, and what we'll do is, we'll still use Optical Character Recognition. We'll use that API to pull the text  from an image, and after that text is pulled, then we already have the Translation API that has been built a few months back. We'll then translate that text, so then it's in the given language of interest. And once we have that translated, then we can pull all of the entities using Name Entity Recognition, and then service and populate all of the fields in the Salesforce object.

So that's exactly how we would go about doing it. This is something that we actually have thought about from the beginning, because a lot of people on my team, and on other sister teams of mine, have told us that internationalization is a huge priority, is very important for Einstein and Salesforce as a whole, and that's something that we should definitely keep in mind, moving forward. We're glad to say that we're definitely working on that right now.

 
Marjan Josenia:
For instant search, as I said, we have two types of queries: conceptual queries and non-conceptual queries. For conceptual queries, currently, we do not cover the non-English words because it needs completely new model to train on the language model of, for example, the Chinese model. But for the key word search, which is non-conceptual queries, we do have [inaudible 00:15:46] that can tokenize the words for us and we can search for that in the Einstein search.
 
Barbara Alberts:
Part of the goal for this podcast too, is just to highlight the diversity of jobs that are available, and the opportunities that exist on just one product, in one cloud. So can you guys both just tell me how these projects have helped make Einstein a better product?
 
Marjan Josenia:
In instant search, when a person wants to find its opportunity, or its slates, it's usually put lots of name entities into the query. So, in order to make a sequel query, which is relevant to the actual query, we need to identify that name entity. And, if it can improve that name entity recognition, we can have a better sequel final query that can retrieve the required information for the user. That's how it helps the instant search.
 
Hanslei Cuoscia:

I'd say that the engine's pretty simple for the business card scanner because that just adds another tool to the tool kit of Einstein, and, especially in the form of a lightning web component, which is a component that can easily be dragged and dropped into a Salesforce display, or a Salesforce grid set-up, where you have your file upload area, where you have your main panel, where you can have your dashboard leads, or any cases, or anything that you're viewing as a Salesforce user, which would be our customers. You could instantly just drag and drop that business card scanning lightning component. That could just be something that they instantly upload files into and get a response from. That's one thing that, I think, it's useful for Einstein for, just expands its tool kit from the existing number of lightning components that they already have.

In addition to that, I think the other project I worked on, which is the Sentiment Analysis Lightning Component, that also adds another really cutting edge and avant garde kind of technology to the Einstein tool belt, which is getting the sentiment for any case or any snippet of text, for our customers. One really useful case that this could be seen in, in the industry, is when a customer wants to look at all of the surveys that have been filled out by their customers, they don't care about the servers that are really good in sentiment or really happy, or really neutral. They want to see the ones that are bad, that are voicing some complaint, or voicing some action for change, because they want to see what needs to be changed, and they want to really act upon that and fix their product, and add a new feature, or remove a feature, to make their product better, and more geared to what their customers preferences are. 

I think the Sentiment Analysis Lightning Component is something that you could easily just drag and drop and select the field that you want to get the sentiment of, will very, very easily allow you to see which case are relevant, and which you should look at, and will save the companies, I guess, thousands of hours. Because they'll only be looking at the complaints, instead of everything, and just filtering manually through what's a complaint, and what's not a complaint.

 
Barbara Alberts:
Can you guys tell me, if at all, if there's any sort of relationship between the work that you guys are doing?
 
Hanslei Cuoscia:
I think one way that Marjan's project could be really helping my team out is that my team is not neccessarily full of any developers that develop these APIs, we simply build on top of them and make them abstractable to a simple one-line of code. From which, you can use that one line anywhere on the Salesforce platform. The code base that we build off of, that code base is developed by Marjan's team, or a lot of people who have very technical depth, and a lot of deep learning in machine learning experience in general. For us, we also have machine learning experience, but not to the point where we're writing a lot of those APIs and all of those lines of code ourselves. But we're understanding what's going on, and we're making it very, very easy for the lay-in and user to just drag around and use anywhere in Salesforce.
 
Barbara Alberts:
To wrap this up, how has your internship helped you in terms of career development, in terms of expanding your research and things like that? How has this summer really helped develop that for you?
 
Marjan Josenia:
Actually, it was very helpful for my research. I was able to read lots of papers regarding the name entity recognition, query parsing, semantic parsing, and I had the opportunity to expand on knowledge in the area that I haven't had any experience before. Also, I had meetings with my managers, with my mentors, and I had the opportunity to get lots of information from them and learn a lot from them. Because their experience in machine learning and DIP learning. They gave me lots of insight when I look at the problem, when I look at the project, and helped me how can I approach a problem that is kind of challenging, and was very helpful for me, actually.
 
Hanslei Cuoscia:

This internship has helped me in countless ways. The first of which is it's the first time I've experienced product management as an internship. Parts of this I was doing program management, parts of that I was purely software stuff. So, this is definitely a new experience and it's been nothing but great. First of all, the mentorship has been great, right? So that's what everyone at Salesforce says. But for me in particular, I really resonate with this point because everyone on my team is very credible, very experienced in this domain. They've learned product for several years, but not only several years, but several years on a machine learning product. So it's definitely very relevant to what I'm working on right now.

In addition to that, they always go out of their way to help me out, even when they have a very, very, very busy schedule. One day, my manager came back from a three-day conference and I was like okay he has no time to help me out. He said I'm very busy, but that same day he ended up staying up with me 'till 7:30pm in the office, and just helping me out on a product bug that was stopping me from making progress into the rest of the week. So I think that was definitely an indication, one of the first few indications that my team is here for me and has my back, unconditionally, and they just want me to succeed. So that feeling of your team just genuinely wanting your success, instead of wanting their product built alone, is very huge right there. I think this has grown my career significantly, and there is no better internship that I've seen.

 
Barbara Alberts:
Thank you guys so much for sharing. It's been absolutely a great time.
 
Hanslei Cuoscia:
Of course. Thank you.
 
Marjan Josenia:
Thank you.