Transcript#

This transcript was generated automatically and may contain errors.

Hey there, welcome to the Paws at Data Science Hangout. I'm Libby Heron, and this is a recording of our weekly community call that happens every Thursday at 12 p.m. U.S. Eastern Time. If you are not joining us live, you miss out on the amazing chat that's going on. So find the link in the description where you can add our call to your calendar and come hang out with the most supportive, friendly, and funny data community you'll ever experience.

We can't wait to see you there. Let's go ahead and get started. I'm so excited to introduce our featured leaders today, Amanda Perez and Romina Mir. Romina is an applied ML scientist and AI architect, and Amanda is a data scientist, and they both work at Central California Alliance for Health. And if I got any of those very long titles and words wrong, they will correct me. Amanda and Romina, it's so nice to have you here. Since we have two of you, I will start with Amanda. Amanda, I would love it if you could introduce yourself. Tell us a little bit about your role and something that you like to do for fun.

Amanda's introduction

Oh, okay. Fantastic. So a little bit about what I do for fun. Well, I've been writing and working with R since 2012, I believe. So as soon as I was able to, as soon as I took that first stats class that required it. So huge fan, always love data. So I would say that currently in what I do as a data scientist at the Alliance, which is what we call it for short, is my day-to-day is in this Posit ecosystem. So I develop in Posit Workbench, and I ship production, Shiny apps, and ML pipelines, automated ETL through Posit Connect. And every, on my whole stack, it's just open source R. Tidyverse is a big part of it. And, you know, connecting with Snowflake underneath, which is our data warehouse.

So that's kind of the day-to-day. Also a lot of workflow optimization, which is where that kind of ETL process that you can really do everything end-to-end in R. You can build an application where folks can go in and do all the work that they need to do in one place, as opposed to, as Romina knows, opening 20 different programs and tabs and scrolling and trying to figure things out for, you know, for one case that they need to work on in a day, you know, so going from a one-hour task to a five-minute task. So that's day-to-day, what I'm doing here at the Alliance.

And what I like to do for fun, I mean, it's in my background, it's in my outfit. Dodger games is an extensive hobby of mine. So it's also family time. I go with my sister-in-law and my niece, and I've been to double-digit games this regular season. So I'll call that my fun hobby. And I'm based in LA, Los Angeles. So I was just gonna say a little more context. Well, you know, it'll take us an hour with traffic to get there, but only a half hour to get back. So it's the story of living in a big city. It's the story. So that's what I'll do for fun. I'll have that be my fun fact.

Romina's introduction

Amazing. Your background is so much fun. Okay. Romina, I would love to hear the same from you, a little bit of background about what you do. Hopefully I got your title right and something you like to do for fun. Yes, sure. So I would say like, actually me and Amanda work very closely together. And we usually focus on internal tooling. So that's the, you know, how we can automate and optimize the, you know, the workflows within the company itself, coming from the company that is less techie for people who are coming from very techie world, I would call us less techie. And for that reason, we are there to, you know, transform the whole workflow for many different groups and different projects and who work on many things.

My work is obviously creating some dashboards, but I think the, I think the lead there is Amanda. Amanda is like great at creating dashboards. Of course, I do as well based on the projects, but my work is mostly focusing on backend, which is mostly machine learning. So I do from supervised machine learnings to all the way to, you know, implementing LLMs, rack systems. And I can obviously talk about more about some of the projects that how that has happened, or how we are using some, you know, external tooling, some open sources to be able to achieve that. But yeah, I think my work is like mostly revolves around some sorts of machine learning. Again, it can be classical machine learning or not. But again, the way it goes out to the staff is usually through the dashboards, but in many of the bigger projects, because we have Amanda who's like perfect at it, I usually like I will hand off my, you know, my models and my, you know, outputs to Amanda to be able to create something. But if not, I'm also doing that on my own. That's how I'm going to talk about my day to day.

Fun fact. Oh, my God, I don't know everything. I think. So my background is a little bit. My let's say my academia background is a little bit different, which is in robotics. And I'm still have some niche of, you know, you know, being still involved in that world. But again, true, true working. My, you know, my working history is always been like health tech companies or health care companies. But yeah, I guess I can talk about a little bit of that. I do. I like robotics. That's what I do or like on my fun time. If not, I'm baking. Both of these things. Oh, that totally counts. And Romina is an AI genius. I feel like you undersold yourself there. You are, I consider you an AI genius. I'm so fascinated. And I mean, the things that you do. Amazing.

Amazing. Thank you. Yeah. Like, again, we really work closely on many of the projects. The projects are what I'm excited to talk about. Well, the way that I learned about Amanda and Romina and that Rachel learned about Amanda and Romina is because Amanda applied for the Posit Impact Awards. And she actually applied twice. So I got to read two of their projects, which are both really, really impactful. It's kind of hard to choose between two really impactful things. But I would love it if you could give us a little bit of a rundown Amanda, and I put a link in the chat as well. So everybody can read about them, about your project in sort of a recap mode. But you have the unique experience of building something that was really going to move the needle. I think a lot of data scientists have the experience of like, oh, I build a lot of things and then nobody uses them or they don't go anywhere. Right?

And this is super common. If this is you and you were building something and you're like, it didn't make this like huge splash that I thought that it would. That's pretty normal. So when we get the rare instance of like, I built something and it actually really, really made a difference. Being able to talk about it and share about it is so great. And also I applaud you for even applying because I hold back. I don't talk about the cool stuff I do all the time. And it's kind of hard career wise for me to do that. Probably hard for other people too. So kudos for that. I would love to hear about the project in a little recap mode and ask you a few questions about that and then we'll hop into Slido.

The provider dispute resolution project

Yeah. But I'll first, I love to touch on your point on building products that folks don't use. So one thing that we love about Posit and Posit Connect is we can track users and usage and on the application. So if there is an application that folks are not using, then we're like, hey, they don't like to really tell you that they don't like it. But we're like, hey, it doesn't hurt our feelings. Why don't you just talk to us and we can make it better for you? So that's why I really love Posit. So Romina, I can just do like the quick first, the first part and then let you talk about your part and split it like that.

So we— Obviously I can talk more of the back end of it. Exactly. It's perfect. Exactly. So Romina and I won the Posit Impact Award for provider dispute resolution claims platform we built, which doesn't sound very exciting, but to us it is. Essentially whenever a provider submits a claim, their claim is denied. They don't agree with that claim being denied. Then it goes to provider disputes and our folks at our company have to work through those disputes. And they were at a huge backlog, Romina, of I believe 7,000 disputes and there are regulations. They're getting fined every day. They wanted to hire 12 full-time employees. And our boss said, you know what? Let Romina and Amanda take a stab at this. And that's how this project was born.

So my part was everything. Romina is going to cover that ML side of it. And then I'm going to cover kind of everything kind of around it that wraps in, makes it look really nice in that dashboard. So the automated data infrastructure that I built for this dashboard, what it did was it consolidated all of the different things that they were doing. I mean, 20 tabs may be an exaggeration, but I know for a fact there are at least six different sources and even on one single source having to click around to so many different places. So I consolidated the cases, the claims, the vendors, the documents, which Romina had a big part in helping with, the notes, work queues. So all of these different systems that they were working into one unified, we like to call it a one-stop shop, that is refreshed daily before the staff log in. And it turned to production Shiny apps on Posit.

So we have a manager dashboard where leadership is able to look at real-time KPIs, SLA, and aging monitor, making sure that they're on track and the backlog is not there anymore, which it's not. They caught up on all 7,000. And then also batch assignment, which is something that Romina worked on. And then this claims viewer dashboard, which is where the workers will, the claim techs will log in. It will recognize their Posit username based on their Posit username. It will show them their cases that they have been assigned. And they've got one place, they click their case, all of the information that they need is on one screen. And it replaced an hour and five minute, I think, was the estimate manual hunt across systems with one single pre-assembled view. And then I would like to hand it off to Romina to talk about all the exciting stuff under the hood, which makes it so powerful for them.

And it replaced an hour and five minute, I think, was the estimate manual hunt across systems with one single pre-assembled view.

Yeah, as Amanda said, before this tool, each case was taking more than an hour. And I think the stakeholders came to us originally, obviously to our manager and like, how can we deal with this? Obviously, I think all of us as somewhat, you know, as data scientists or, you know, with the data scientist background, we know that like, usually people who are not in this world, they don't know what they don't know. They didn't know how to handle this at all. And so we had to look into how we can like triage this workflow. So one of the main things I try to do, I try to see, first of all, is there is there a there there in the data can, because, for instance, like they taught, all of the cases are different. There is no way that we can predict a case, like we cannot predict an outcome by any means. And I was like, Okay, let's see what the data says. I know you have to look into like, I don't know, 20 data sources. But let's see if I only pick even the 15 that is like easiest for me to catch. Looking at these 15, you know, data points, can I have something to showcase? And the answer was yes.

We tried to actually look into the accuracy and precision. So not just accuracy, because it's really important, like, where the, you know, false positives are, again, we're dealing with claims, and we're talking about people's and, you know, we couldn't just like, close our, close our eyes on something, right. So one of the main things I did was be able to triage and be able to look into the triaging, the outcome, and looking to, okay, for the most outcome that we have, can be and like triaging it based on like, how probable it is to get there. And based on that, who can work on it. So that was the main, like one of the main things we tried to do. But I think it's important to mention that, like, usually, as data scientists, I think we need to be able to come up with different ways. So when we did that, we thought it's done, like we thought, okay, we are already giving you the probable answers, like final outcome of each of these cases, it should be enough for you to be able to, you know, go and ask for the staff to go and find their way and do your work. Apparently, that was not the case. So we had to see, okay, okay, how can we again, we brought the question back in, we said, okay, while we while we added one layer that was helpful for maybe the leadership, and perhaps the staff, the staff is not using it as it should be, as it could be per se, even though we're giving them all the sharp values, all the information that they can leverage. So we brought the question in, we tried to look at it differently.

One of the things we realized is that previously, the managers and the leadership, they didn't know

whatsoever how one like, how many how many cases of staff is working on staffs were blind of how many cases it existed. And we said, Okay, that's probably an issue that at most what the managers were doing, they're using Excels, they're using like exporting all the information on a daily or weekly basis, assigning people and everyone should like, go from Excel to another tab and from that tab to another 20 tabs. And we said, Okay, that's definitely not helpful. So I tried to bring in and based on the skill set. So I look at the skill taxonomy of the staff, obviously, with the help of the, you know, stakeholders. I tried to look at the skill sets, so skill taxonomy, and basically assign auto assign cases to people. Obviously, managers at this point can can say, Okay, maybe this week, I want, I don't want these people to work on these cases, or I don't want to, I want to delegate cases differently. And they can still as well. But now true managers dashboard, which is again, one of the main things I'm under worked on with the auto assignments, the manager knows who is getting how many cases, how long is taking them to work on something because in the past, it was again, it was just like a black box. No one knew, they had an estimate of like, okay, it's taking everyone more than an hour, I think an hour and five minutes to be able to close the case. And then now they know who is faster, who's who's not closing enough cases. And so they can have a look a closer look at their own stuff as well. So I think that was the second machine learning piece I worked on.

And a smaller piece was the PDFs to be able to integrate bring all the PDFs in all the documents, let me not call them PDFs, like all the documents seen, because some of them are Excel, some of them are as PNG, some of them are PDFs. And some of them were dot tips. That was your biggest issue that you had to take the dot tips, yeah, hurt them and then have them read it. So I mean, that was yes. So that was again, that was another thing we added, because yeah, so I guess that's the full story, like from end to end where we started. But again, it was an iteration, it was an iteration of couple layers and adding more and more to it to be able to call it a final product. That resulted in now talking almost zero past dues and whoever works with claims, they know like, how difficult it is to deal with the claims. And imagine when you have like past due claims, you're now being charged, and so on, so forth. So we kind of like mitigated that.

And I know that you had, you probably had an SLA of like 30 days, right? Where you were like, we need to get these done in a certain amount of time. And yeah, yeah, backlog, depending on contracted or non contracted 30 or 45 days. Continually missing SLA, which makes leadership really, really nervous. I think that the, the thing that I loved was that you didn't have to hire those 12 or 13 extra people, right? Like you were looking at, okay, we're going to need to double our team size, but you don't need to now. So now you have 12 people who can do the work of 45 people when that sort of workload exploded.

So now you have 12 people who can do the work of 45 people when that sort of workload exploded.

Do you think that the workload exploded because of AI on the other end? Like AI made it easier for people to submit the disputes. And then that's what caused the number of disputes to get bigger on your end? I don't think so, because the provider portal is also not a great system. It's a little antiquated. We've not gotten to touch that or help with that yet. So I think it really was the fact that they were spending over an hour on a single case and they're getting hundreds of cases a month. And they just stack up.

And then with Romina's auto assignment, the other issue is that certain claim checks can only work certain types of cases. So if there's like a DME, a durable medical equipment, you have to be trained in how to work on that kind of dispute. So I feel like it was really just like an oversight, like a managerial aspect of, they were just doing the best with what they had and they were drowning. And they just thought adding more people would solve the problem, which is why you need a better tool. You have to kind of think beyond what they think that they need. And I think Romina touched upon that, because sometimes they don't know the possibilities. They can't see the vision. And now, especially this manager view that we have, our COO is on that like every single day, just being able to have the ability.

Yeah, it was a huge problem for the company to the point that like everyone was coming in to see how they can stop the bleeding, basically. And well, at some point again, because we brought, we've been brought in, we said, wait, wait, wait, I understand you want to hire more people, but let's see what we can do in the interim. And like we tried to, and I think one of the, at least internally, I can talk about our team. I don't know other teams' workflow, but like we have like a very quick turnaround. We believe in iterations. We don't want to look into the, like giving a perfect answer to the stakeholder. We're going to do iteration. So we built something very quick, something workable, reliable for sure, but like workable, very quick, even on a draft version, we send it over, we get feedback, people are working on it. And while stakeholders are giving us some feedback, we're going to go and go to the second iteration and the third iteration. And that's how the first machine learning was implemented. And the second machine learning was implemented. So it was like a very hand in hand and very close work to be able to achieve what we achieved. But I think one of the good things is we didn't want it to have the perfect answer in the very beginning because we didn't know what the perfect answer is even, because again, I taught triaging with the, you know, probability of the outcomes. That would be the answer, but probably that wasn't the bottleneck necessarily. Maybe the bottleneck was like having some visibility of how staffs are dealing with each case perhaps. So again, some of the things we don't know, probably stakeholders might not know either, sadly. So we have to like work together and learn this together. Yeah.

Audience Q&A

It's a conversation, communication. All right. Well, we have so many questions in Slido. If we do not start asking them, we will never get through any of them. Let's do it. Yeah. Romina, I get so excited talking about this. No, it's been wonderful. And we have a question. In fact, the one that has bubbled to the top, because reminder, everybody, you can upload questions, is right in line with what Romina was talking about a little bit, because the question says, in your experience, what are the key indicators that a problem is a good candidate for machine learning rather than a simpler statistical approach? And that sort of iterative process is probably part of that. Who wants to take this one? Is this a Romina question?

I can take that one. I think it's about what's the question we are answering. So here we did wanted to have a more predictive workflow instead of just like, because statistically, if we want to only use a statistical point of view, you're only looking at historic data. You're not looking at the future information. And I think what at least we wanted to influence, it was like maybe some predictive analysis would be an answer. And I think that's the way we kind of took it on this case. Again, we touch base many, many, many, many things. And for instance, Amanda does a lot of ROIs and causal analysis and more of a stat side. But because we're now focusing on the impact award and the project we did, which is, again, one of the biggest ones throughout the company as well, this is the approach we took. I hope I answered the question. I'll be happy to answer any follow-ups. There are plenty more questions for that case. Do not worry. Hubert had a question. And if he could ask that live, that would be amazing.

Oh, my computer is telling me I need to update. Hold on. I'm ignoring that. Sorry. That was my being slow there. There we go. There I am. Yeah. So this is really interesting. And I think my question is, what kind of data sources do you use? You talk about claims data, EHR data, other kind of data, you've got provider-based data, or some of those, I mean, you've already talked some of those PDFs and TIFFs and all the crazy data challenges, but what are those data challenges, some of those tough data challenges that you've had and how have you solved them? And is it data challenges in terms of availability or structure of data or both? Everything. That's a great question, because that's what I struggle with since the first day I started here. So when I first started, initially, I thought that the main data issue was going to be the fact that we primarily rely on claims data, which has about, it can have up to a three-month lag before we can say, okay, we've, it's been enough time. If a claim happened, it would have been in our system. Now they have that amount of time. So we always have to take into account a claims lag when we do our analyses. But then as I started getting into more ROIs, program evals, building things, it was actually the sources of data, the quality of the data, not knowing the lineage of the data, having to ask like 20 people, hey, did you make this table? How did you make this table? Where's the code? So I kind of had to do my detective work to get to the top to find what are our sources of truth. And I feel like that's a big struggle because there are, we do have a lot of different vendors.

So for example, for, for the, this project and Metatrack, which is a vendor, but in a way that's good because then we get source of truth, reliable data from them every night. So we can pull directly from those tables. On other projects that we're working on, they use a vendor Jiva for their UI and management. So then we're able to pull directly that Jiva data. The, the, the tricky part is the data storage, the data, you know, there's no oversight on folks making tables. So it can be a little bit tricky, but my solution where now I feel good about it, and I don't have a heart attack is that I'm making sure that when I'm pulling in data, I'm like, I'm from, I know where that data came from. I know that it's a direct ingestion from a vendor. I know that it's coming directly from a state file. And finding the locations also can be difficult. Romina can also speak to that. I mean, with the PDF files or the TIFF files, they are stored in some insane locations. So it really is, I think, from an organizational structure, the data governance, maybe we want to call it, it's just not where I would love it to be. I mean, I come from academia where, you know, data and gathering data and storing data was a lot simpler. So it was a huge, it was a huge challenge. So for me, I just have to always ask myself, what is this, where did this table come from? Is this the original ingestion? Because I don't want a table that somebody else has, has messed with, which has, has been the biggest headache.

Amazing question, Hubert, thank you for taking us in a good direction. Lots of conversation in the chat about, in data quality or whether or not third-party data is reliable at all. And I want to make sure as well that we define TIFF out loud. I define it in the chat, but a .TIFF file is just an image file, which is a difficult thing to parse. Yeah. When it's full of information, PDFs are hard enough, but image files are notoriously, notoriously difficult when you have them as your sources of input.

Okay. So we have some more questions here. There is one that I would love to ask both of you just to get a little bit of context about the way that we work here. Wait, I lost it because things moved around in Slido. Aha, it was Zach's question. I have put it to the top here and I will put it in the chat as well. It's got a little asterisk, so I'm going to read it for Zach. And Zach says, after automating a lot of processes, do you find yourself with not much work to do? And if so, what would you do in that kind of situation? I don't know if this is ever a problem for you. There's probably more and more stuff to fix, right? Like you said, there was like a portal that you hadn't even gotten your hands on yet, right?

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r

Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r Amanda Perez & Romina Mir & Data Science Hangout, shiny-r