<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Paola Neira on Medium]]></title>
        <description><![CDATA[Stories by Paola Neira on Medium]]></description>
        <link>https://medium.com/@paolaneira?source=rss-2693af22b2aa------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*kgDKoJHyBAWsivxkxfUqUw.jpeg</url>
            <title>Stories by Paola Neira on Medium</title>
            <link>https://medium.com/@paolaneira?source=rss-2693af22b2aa------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 02:44:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@paolaneira/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Meet Vic —Engineering Lead of the Predictions Squad]]></title>
            <link>https://engineering.rappi.com/meet-vic-tech-lead-of-the-predictions-squad-48adcb4d8c9f?source=rss-2693af22b2aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/48adcb4d8c9f</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[rappi]]></category>
            <category><![CDATA[interview]]></category>
            <category><![CDATA[engineering-mangement]]></category>
            <category><![CDATA[predictions]]></category>
            <dc:creator><![CDATA[Paola Neira]]></dc:creator>
            <pubDate>Fri, 27 Aug 2021 17:31:23 GMT</pubDate>
            <atom:updated>2021-09-07T22:15:21.600Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>A chat with <a href="https://medium.com/u/85bfd53e8edb">Victor Hugo Camargo</a>, our Engineering Lead of the Predictions Squad in the Fulfillment Area.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rT0lrgNr6ZYCYt2UwucNDw.png" /></figure><p><strong>What do you do?</strong></p><p>I am the technical lead of a team full of exceptional people with great abilities. Our job is to take operative challenges and turn our ideas into engineering products, that impact positively Rappi’s operation.</p><p>The predictions team is in charge of providing the operation with two things: i) time estimations (ETAs) and ii) supply and demand forecasting used by all the operations.</p><p>My job is to deal with diverse and interesting ideas, from how to implement features to take our models to the next level, all the way to how to implement fast and scalable architectures capable of serving millions of predictions per minute.</p><p><strong>How did you get into engineering?</strong></p><p>I’ve always had a huge passion for technology; I’m always trying to understand how things work and why they do what they do. Because of that, I major in physics engineering, as it equipped me with the fundamentals that would enable the understanding of the world around me.</p><p>Throughout my career, I was always surrounded by problems that encourage me to learn a little bit more; challenging me to expand my knowledge over time. I started building plasma reactors, developing communication tools, robots that deactivated anti-personnel mines, among others. Fast forward, I now lead a Machine Learning team. Having the opportunity to solve very interesting problems, with advanced technology in the ML world.</p><p><strong>What inspires you the most?</strong></p><p>The impact Rappi has over millions of lives of users, couriers, and partners. We offer our clients magic while facilitating their lives by offering a large variety of products and services in record time. All of these while being able to offer couriers to get an additional source of income to bring home. Every single day we work so that we can make couriers more efficient in every order thereby increasing their productivity. Ultimately, we are optimizing every single thing within the process to give all of our different users more time.</p><p><strong>What’s the best advice you’ve ever received?</strong></p><p>“Have a coffee and re-think it”. It does not matter how brilliant the solution is, or how difficult the problem seems. Rethinking the problem with a coffee makes you look at the problems from a different perspective and by doing so, you will be able to find something you didn’t notice before. It’s always good to iterate the ideas with a coffee on hand whether it is alone or with a team member.</p><p><strong>What advice would you give an engineer that’s just starting out?</strong></p><p>“Be gentle and ask away”, you will always find experts in certain topics while you’ll get people less knowledgeable in some others no matter the complexity of the subject. In a few years, you will be the sum of all the answers given to you by people you’ve interacted with, and of those ideas that you’ve given back to others. Be gentle and curious.</p><p><strong>What’s your favorite part of the day?</strong></p><p>Coffee time, would you guess? That’s when I get to discover and discuss ideas with different members of the team. It’s the time of the day when magic happens.</p><p><strong>What’s the most challenging or exciting project you’ve tackled at Rappi?</strong></p><p>Definitely, Rappi-maps architecture has been very challenging. Basically, Maps is our in-house solution for travel distance and time estimations used by the company. It receives +20K requests per minute on an average day. It’s a project with an immense impact in terms of cost savings for the company but also in terms of benefits for the operation.</p><p>Rappi-Maps is a problem that evolves with the size and necessities of the company, as we grow. Right now we’ve achieved a large improvement that has managed to reduce the resources used by the microservice by more than 90% which has brought us to a point of stability.</p><p><strong>What was the first thing you ever develop?</strong></p><p>As a good geek I love board games, and as a curious student, I always wanted to know what would happen if you would have played perfectly, calculating every single probability of things that could happen in the game.</p><p>My first coding project was a risk simulator where I could try different ways to play games. I was 13 years old, had a good C++ book and plenty of time to work on it. This project stayed with me for a while, I translated it into Java, Python, Ruby y JS. Honestly, it’s not something out of this world but it helped me to think about how to solve problems through software. It became my personal “Hello World” and it’s the first challenge I use while dictating lessons.</p><p><strong>What is something you’re most proud of in your life or career?</strong></p><p>Working in the dispatching team at Rappi. We worked hard, slept little, and got to learn an infinite amount of stuff that made us very happy. We managed to assign orders in a very efficient way and every single improvement we did was immediately translated into minutes for the operation, which would ultimately result in more happiness for all of our clients.</p><p><strong>What role do dogs play in your life?</strong></p><p>Maxwell is my best friend. He’s always there for me to remind me that I need to catch a few rays and that life is better when you run behind a ball.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hVPCvSiWX6F1XbUym1OH-Q.png" /></figure><p><em>Vic always thinks better with coffee.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=48adcb4d8c9f" width="1" height="1" alt=""><hr><p><a href="https://engineering.rappi.com/meet-vic-tech-lead-of-the-predictions-squad-48adcb4d8c9f">Meet Vic —Engineering Lead of the Predictions Squad</a> was originally published in <a href="https://engineering.rappi.com/">Rappi Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding Lost Demand]]></title>
            <link>https://engineering.rappi.com/understanding-lost-demand-1088f8b4dc92?source=rss-2693af22b2aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/1088f8b4dc92</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[logistics]]></category>
            <category><![CDATA[simulation]]></category>
            <category><![CDATA[rappi]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Paola Neira]]></dc:creator>
            <pubDate>Wed, 11 Aug 2021 20:50:40 GMT</pubDate>
            <atom:updated>2021-10-20T02:15:22.699Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*48UhYA613nnCQzMyO8heVA.png" /></figure><h4>What does lost demand even mean?</h4><p>It’s Saturday night in Sao Paulo, the governor has established new lockdown measures just yesterday and a typical Brazilian rainfall starts draining the city. What starts happening from that point onwards is one of the most complex problems we face at Rappi.</p><p>As it continues to rain, Brazilian stomachs start feeling like <em>pao de queijo</em>, and as a result plenty of users start opening the app. A few minutes later, orders start coming in at a fast rate and our logistics engine starts distributing orders across our partners. Unfortunately, we quickly realize that there are not enough couriers connected to match supply and demand. Not to mention, a portion of the fleet is actually disconnecting due to the rain. Seems like we could have known this before, but unfortunately the weather forecast for the day was giving us a bright and sunny day in ciudad Paulista, these are the perks that the equatorial region gives us 🙃.</p><p>If we have limited supply availability relative to customer demand, we may need to remove fulfillment options (“levers”) from the app to avoid overcommitment. Users at certain zones would face higher ETAs, or coverage reductions in favor of system efficiencies. However, when this happens, some users would rather cook at home, pick up the food themselves or just grab a bowl of cereal instead. These are examples of what we call in the industry “lost demand”.</p><p>Measuring lost demand has been one of the most complex problems we face in the on-demand world. Especially, as our system combines the unpredictability of the equatorial region with the multi-verticality of our value proposition and the different transport types used by the couriers. Regardless, understanding lost demand is key in order to create a virtuous cycle as it will end up being part of the further ahead forecasting process. If we fail to measure it, we are setting ourselves up for failure as we would underestimate our demand, calculating less required couriers connected which will end up dampening demand again, starting the cycle over.</p><p>Ultimately, the rationale of what users do in the app is totally unknown for us. There are multiple reasons for users to stop engaging the app before they place an order: they might have received a text from their neighbor inviting them for dinner, disliked the cuisine availability at that point, perceived ETAs as very high, or were just food browsing to get ideas on what to cook. So, in order to understand how our demand would have behaved if no levers would have been applied, we must evaluate the following scenarios.</p><h4>Approaching the problem</h4><p>At first, we thought the problem could be solved by using a regression variance analysis, something like the below to state it simply:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jLhZw64V5T4zKr1T" /></figure><p>That approach would allow us to estimate the effect of different actions (factors in statistical language) over the dependent variables. We could isolate the specific treatments and factor them in our analysis to understand each of the levers on its own and the multiple combinations or interactions between them.</p><p>However, we encountered several pitfalls in the process; i) we were facing the so-called network effect, which you could find best explained <a href="https://tech.instacart.com/it-all-depends-4bb7b22e854b">here</a>, and ii) establishing a baseline scenario became challenging as the data showed what happened historically, and the operation applies levers to demand always under these scenarios. In short, our data was already biased. As a result, we decided not to go with regression analysis as it could only estimate a small portion of the components or interactions within the system. We were back on square 1, and we needed to add a new layer of complexity. On top of estimating lost demand, we wanted to understand if there was an opportunity to lose less without compromising service. Ultimately, we were trying to understand the tradeoff under different scenarios so that we could find an optimal point where we could maximize conversions while minimizing the probability of a defective or cancelled order.</p><p>As a second option, we thought about using our experimentation tools, which allowed us to test hypotheses and analyze data, resulting in fast and reliable insights. We tried different frameworks such as random switch or factorial design. However, we could not isolate the effects described earlier either, and we kept facing high interdependency between the entities in our system and within the entities and the levers applied. On top of that, testing in production meant we could be risking some users’ experiences. We needed something more robust and risk-free to add value to the business.</p><p>For our final solution, we opted for a simulation tool, which we decided to call SimOn, short for “online simulator”. SimOn runs 10 hours of production in 10 minutes given its combination of fast data processing and an event architecture. The main framework to do simulations is a combination of discrete event simulation with high-frequency simulation steps and some agent-based modeling.</p><p>To model out the operations, we established actors for every entity in our environment and translated it into a simulated world; we re-created the interactions between them and their decision-making process (“policies”). Some of the policies were modeled using stochastic processes, in which we used fitted distributions from historic/input data, as we know that these policies have uncertainty and a certain independence from the other processes, like the inter-arrival time, which is used to model the order velocity (or rate of order creation). Others were done through machine learning, like movement time estimations or preparation time of an order at a restaurant, which needed “contextual” variables to get rough estimates of the processes.</p><p><strong>With all the machine learning hype, we frequently forget how other approaches can be of greater use in specific scenarios where a prediction is not enough. In our particular case, we needed to estimate how our system would behave given specific characteristics and model the specific relationships and decision-making processes for our actors (users, restaurants, couriers, and Rappi) to test our hypotheses. We also wanted to do this under multiple scenarios to understand what the sweet spot between both growth and service was. First, we built a base simulation mode, in which we model the exact historical data to compare the aggregated metrics with the input data; this enables us to debug and measure the level of certainty of the simulator. At last, we built a demand generator using bootstrapping and some sampling strategies from historical curves of both phenomenons and past characteristics from data and masking some of the attributes for the timestamps of our interest; from that process, we get a hypothetical set of orders to simulate in SimOn.</strong></p><p>We are currently in the process of testing out SimOn to find if it will ultimately allow our data scientists and operations research engineers to create different demand scenarios with varying combinations of balancing levers, for example, increasing or decreasing our partners’ availability given a certain number of orders being prepared at their restaurants. The latter so that we unleash questions like: “What if we double our demand in these specific zones in a particular city? Or what if we are more relaxed with our balancing levers? Do the service metrics get worse when we apply x or y policies? Can we stress the system more and keep operational metrics at acceptable levels?</p><p>As mentioned earlier, when we close out fulfillment options, we are trying to avoid compromising the service of our users but we definitely do not want to underestimate that ability, so enabling tools to crack this sweet spot are of incredible value for our business.</p><p>The state of the art of the logistics discipline is continually being created and as a result we keep looking for new paradigms to increase our system efficiency and continue to add value in the beautiful but unpredictable Latin America. We’ll keep you posted on SimOn’s further achievements and pivots.</p><p>We would like to thank all the engineers who helped us build SimOn.</p><p>Authors:</p><p><a href="https://medium.com/u/c18a1119ebbf">Alejandro Cadavid Romero</a></p><p><a href="https://medium.com/u/2693af22b2aa">Paola Neira</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1088f8b4dc92" width="1" height="1" alt=""><hr><p><a href="https://engineering.rappi.com/understanding-lost-demand-1088f8b4dc92">Understanding Lost Demand</a> was originally published in <a href="https://engineering.rappi.com/">Rappi Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Opening the Pandora’s Box of Machine Learning]]></title>
            <link>https://medium.com/@paolaneira/opening-the-pandoras-box-of-machine-learning-4cdf7a80a043?source=rss-2693af22b2aa------2</link>
            <guid isPermaLink="false">https://medium.com/p/4cdf7a80a043</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Paola Neira]]></dc:creator>
            <pubDate>Mon, 26 Nov 2018 18:44:12 GMT</pubDate>
            <atom:updated>2018-11-29T16:35:45.972Z</atom:updated>
            <content:encoded><![CDATA[<h4>A short and easy overview of Machine Learning…</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0pX5JMZtVTBo7rR1yu0CnQ.gif" /></figure><p>Machine Learning (“ML”) is shaping every aspect of our life and it will continue to do so as continuous discovery and innovation keep happening in this field. As a result, understanding the intuition behind it is crucial to avoid waking up in a world full of technology that feels more like magic than like science. It’s true that Deep Learning seems like a black box where any data can be thrown in and enlightening insights come out, but understanding what happens in that box will allow you to truly appreciate its power but also to realize its challenges and limitations. This mini-series of posts have the goal of quickly getting you up-to-speed with ML through the following chapters:</p><p>1. Definition of Machine Learning.</p><p>2. Types of Machine Learning.</p><p>3. Deep vs. Shallow Learning.</p><p>4. Applications and when to use which type of ML.</p><p>5. Challenges and Limitations.</p><p>6. Getting started (even without knowing python).</p><h3><strong>Part I: What is Machine Learning? Is it the same thing as AI?</strong></h3><p>Most of the times we see the words Machine Learning and Artificial Intelligence (“AI”) seen as interchangeable terms, but Machine Learning is in fact a sub-segment of Artificial Intelligence. AI is the broader concept of machines simulating any task considered “intellectual”, such as reasoning and knowledge representation (Watson on Jeopardy), ability to move and manipulate objects (robotics), and natural language processing (speech recognition), among others. Weak AI refers to the case in which a machine is solving a specific problem whereas strong AI or Artificial General Intelligence refers to the case where a machine has the ability to apply intelligence to any problem (siri). Just like machines simulated the body in the industrial revolution, AI is looking to mimic the brain.</p><p>What is considered an intellectual task then? For the purpose of AI, it is a moving definition, and it is defined by every success in the field. Once one achievement goes mainstream, it is no longer considered AI. This is called the AI effect and is the reason why we no longer consider a computer playing chess AI.</p><p>What is Machine Learning then? It refers to the practice of achieving AI by giving machines access to data and let them learn by themselves. In other words, using given data, the computer will recognize a pattern in order to make a prediction, which will then use to optimize the error/loss function; this last step is the “learn” in machine learning. To further build on this process let’s use a very simple example to illustrate the process. If we were to estimate the price of apartments in New York using features such as area, neighborhoods, number of bedrooms, number of bathrooms etc., we could think of the model below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/982/1*pnTWn96LWUkiUjcuN79MJg.png" /><figcaption>Linear regression determining price of apartments in NYC</figcaption></figure><p>The process runs like this: First, multiple data points of apartments’ prices and features will be inputted to the algorithm. Second, based on this data, the algorithm will recognize a pattern (if it were a linear pattern, the algorithm will have a model just like the one above) and make predictions for the coefficients next to the variables (the βs in the equation). Third, it will take one of the data points inputted and estimate the price based on the coefficients calculated and the features from the observable data points. Fourth, it calculates the difference between the prediction it did and the one in the observable data. Finally it will take that difference (the error)and use it as an input in its future predictions. It will run this process over and over again, until it reaches the point where it has minimized that error function.</p><p>Sounds familiar right? Just like linear regression in statistics! Well yes, Machine Learning and Statistics are highly related. Just like statistics in ML, the regression doesn’t need to be linear, it can follow any type of function. However there is one important difference between them, Machine learning requires no prior assumptions about the underlying relationships between the variables, which in practice means it does not set the data to fit a certain model taking into account statistical properties of the estimator (p-value, unbiased estimators), the underlying distribution of the population, and the expected properties, it is instead like a black box that just sets the model based on the inputs it receives and therefore the more data you have, the more accurate your prediction is. This is where Big Data and Data Scientists come in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/765/1*js1aU9F9zuHWeS5O-NH29A.png" /></figure><p>Data science is different from AI and ML because its goal is a human one: obtaining insights from the data, including descriptive, explanatory or causal. In other words it refers to the ability to take data, to be able to understand it, to process it, to extract value from it, to visualize it, and to communicate it. An ability that comes quite handy in a world in which there is such a huge amount of data. So, is finding the growth rate of a company’s revenues in excel Data Science? No, not everything that produces insights qualifies as data science. In fact, the classic definition of data science is that it involves a combination of statistics, software engineering, and domain expertise.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4cdf7a80a043" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>