<?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 Ariana Bray on Medium]]></title>
        <description><![CDATA[Stories by Ariana Bray on Medium]]></description>
        <link>https://medium.com/@aribray?source=rss-2783dcac0d93------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*9yAb4VME2bEC841JoAsTIg.jpeg</url>
            <title>Stories by Ariana Bray on Medium</title>
            <link>https://medium.com/@aribray?source=rss-2783dcac0d93------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 20:52:46 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aribray/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[Machine Learning For All: An Introduction to Nordstrom’s Machine Learning Platform]]></title>
            <link>https://medium.com/@aribray/machine-learning-for-all-an-introduction-to-nordstroms-machine-learning-platform-9b077b7f212b?source=rss-2783dcac0d93------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b077b7f212b</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[platform-engineering]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Ariana Bray]]></dc:creator>
            <pubDate>Fri, 17 Apr 2020 15:31:01 GMT</pubDate>
            <atom:updated>2020-04-17T15:31:01.364Z</atom:updated>
            <content:encoded><![CDATA[<h3>ML For All: An Introduction to Nordstrom’s ML Platform</h3><p><em>Our strategy for enabling self-service machine learning at scale</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VMoVV38AbkhzzMgN" /><figcaption>Photo by <a href="https://unsplash.com/@pietrozj?utm_source=medium&amp;utm_medium=referral">Pietro Jeng</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>From managing supply chain logistics to <a href="https://medium.com/tech-at-nordstrom/ranking-looks-at-nordstrom-using-machine-learning-b62d0b8b42b3">building curated Looks (digital styling recommendations) for millions of customers</a>, more and more teams at Nordstrom are harnessing the power of machine learning (ML) to continue to enhance the customer experience. From picking a data source to choosing a programming language, everyone does it differently. Where does Nordstrom’s Machine Learning Platform team come in? Introducing <strong>Cerebro</strong>, a machine learning platform for everyone. In this article, we’ll discuss our vision for Cerebro, the need for a machine learning platform, and the problems that we addressed when designing Cerebro.</p><h3>Our Vision</h3><p>Our vision for Cerebro was to create a platform that empowers data scientists, analysts, and other ML developers to create scalable ML solutions that drive direct business impact — in other words, our mission is to enable <em>self-service machine learning at scale. </em>Cerebro aims to provide users with a unified set of tools they can use to rapidly develop and confidently deploy their ML solutions.</p><p>In summary, the main purpose of Cerebro is to provide an easy way to manage a model’s lifecycle. Cerebro directly impacts all of Nordstrom not only by allowing models to be pushed to production faster, but also by allowing other teams to use the Cerebro SDK to replace their services with our ML solutions. Cerebro has already been used by multiple teams within Nordstrom to serve their projects. Our platform has been used to power internal tools such as model forecasting. We keep the following guidelines at the forefront of our design decisions:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/801/1*JUVkfwcv4V2OyQYSQaxCZg.png" /></figure><p>We accomplish these goals by using automations to create a comfortable user experience.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dY9bPRIE7mjSJemDKnO9-w.png" /><figcaption>Automations for a comfortable user experience</figcaption></figure><h3>The Need</h3><p>For a long time, Nordstrom data scientists have looked for new ways to solve this question—<em>how to enable a production-grade end-to-end machine learning workflow</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hL9nHVcRFK2ErJjpobEW5A.png" /></figure><p>The workflow also needs to be inclusive of the many aspects of machine learning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w3GCZ1YUW-QUO9HLNOQKHw.png" /></figure><p>Embedding data engineers within each team to create one-off solutions for these needs simply isn’t a cost-effective option. Allowing those data science teams to manage the whole machine learning lifecycle (from data ingestion to model serving) would take an unacceptable amount of time, slowing down the organization as a whole. In short, we needed a way to <em>abstract engineering from machine learning.</em></p><h3>The Opportunities</h3><p>There are <a href="https://www.xenonstack.com/blog/mlops/">widely known problems</a> that come with deploying a model to production. In addition to those problems, the ML Platform team also has to tackle Nordstrom-specific problems:</p><ul><li><strong>Many Models, Many Standards</strong>: Nordstrom uses enterprise-grade metrics and logging solutions — but without one point of contact for the results of those metrics and logs, how truly effective are they? Who is responsible for maintaining the health of the service and the load balancer? What are some metrics that should be monitored?</li><li><strong>Scale</strong>: Data science use cases are limited to a single, standalone compute instance. Because of the absence of an enterprise-grade ML platform, distributed compute is not leveraged.</li><li><strong>Workflow</strong>: In the absence of ML workflows, engineering teams at Nordstrom often have to make significant adjustments to a data science-led project to make it production-ready.</li><li><strong>Learning New Tools: </strong>In the absence of an abstraction between ML and engineering, those working on ML projects must learn new tools and technologies. This effort is time-consuming, requires engineering support, and adds delays to the timeline of creating a production-ready model.</li></ul><h3>Why Not Just Use an Existing Platform?</h3><p>There are many gaps in existing ML SaaS offerings:</p><ul><li>No centralized feature store</li><li>No support for all models and languages</li><li>Software-specific models</li><li>No model tracking; no <strong>code →feature → model</strong> lineage</li><li>Framework-specific, generalized hyperparameter tuning</li><li>Integration challenges with existing infrastructure</li><li>Costly subscription models</li></ul><p>We decided to build a platform that gives data scientists what they need <em>and</em> fills many of the gaps left by other ML platform solutions.</p><h3>Why We’re Building Cerebro</h3><p>Cerebro is a platform that democratizes machine learning. At its core, Cerebro strengthens the partnership and bridges the gap between data science and engineering.</p><p>Cerebro offers users a range of benefits, some of which are highlighted below:</p><blockquote>* <strong>Self-service</strong> — the user’s reliance on a separate engineering team is lessened;</blockquote><blockquote>* <strong>Consistent </strong>— our approach aligns with existing ML workflows used by the data science community;</blockquote><blockquote>* <strong>Fewer New Technologies — </strong>the need for a user to learn new technologies is reduced to the absolute minimum because we provide abstraction through our interfaces;</blockquote><blockquote>* <strong>Easy to Use — </strong>Cerebro does not slow down the process of <a href="https://towardsdatascience.com/what-does-it-mean-to-productionize-data-science-82e2e78f044c">productionizing</a> models.</blockquote><p>Rather than hiring more engineers to support ever-expanding data science use cases, we have built a <em>platform</em> that can scale to enable those use cases. Instead of making third-party software-driven decisions, we want to make custom algorithm-driven decisions.</p><h3>Benefits of Using Cerebro</h3><h4>Come as You Are: Abstractions for a Standardized Workflow</h4><p>Our team didn’t want to take a prescriptive approach and ask every team to use the same tools and technologies. Instead, we’ve developed abstractions that help create a standardized workflow — the calm <em>over</em> the storm. Users have an assortment of ways to create, develop, and deploy their projects.</p><h4><strong>No Proprietary Technology</strong></h4><p>We are not introducing any technology that’s proprietary to any company or organization — therefore, we have no licensing costs.</p><h4><strong>No New Compute Infrastructure Needed</strong></h4><p>We are leveraging the compute platforms already established at Nordstrom, not setting up any new ones.</p><h4>Choose Your Own Data Source</h4><p>Rather than ask every team to use a specific data source, Cerebro allows users to securely connect to several of the most popular databases and data warehouses. As the platform matures, we will enable connections to even more data sources.</p><h4>Cloud Provider-Agnostic</h4><p>While Cerebro does have some dependencies on specific cloud providers, we are building towards a fully cloud-agnostic solution which can be easily extended to support multiple cloud offerings.</p><h4><strong>No Restrictions on Model Training Frameworks Supported</strong></h4><p>Cerebro supports most of the popular ML frameworks out of the box. Any framework a data scientist would want to use can be supported with minimal-to-no-changes.</p><h4><strong>Lineage Between Code, Data, and Model</strong></h4><p>With valuable metadata available at every step, Cerebro provides a way to tie together the <strong>code</strong> written → the <strong>data</strong> used → and the generated <strong>model</strong>. Users can gain information about how the model is served and also enable model performance/effectiveness measurement.</p><h4><strong>Testing and Quality Control</strong></h4><p>Cerebro makes it easy to seamlessly test logic as part of the Continuous Integration process. The quality of the data is asserted at every step.</p><h4>One-Stop Shop</h4><p>Cerebro can handle it all. Users can:</p><ul><li>create a machine learning project with its own Gitlab repository</li><li>perform local development</li><li>serve models offline and online</li><li>orchestrate the project with Kubernetes or Apache Spark on Kubernetes</li><li>manage project pipelines with enterprise-grade, open source software</li></ul><h4>A Tale of Two Users</h4><p>If a user doesn’t care for the abstractions, they don’t have to use them. However, the vast majority of users enjoy using our abstractions to complete their projects.</p><blockquote>As the Machine Learning Platform team continues to build Cerebro, the principles to which we hold ourselves serve as important reminders of our end goal.</blockquote><h4>Our Principles</h4><ul><li><strong>Modularity</strong>: Cerebro is a lot more than just an endpoint that data scientists pass models into — we have developed an SDK and have multiple showcase projects so that data scientists can see the best use-case for their projects. By enforcing best practices within our SDK, our users can <strong>independently</strong> use our platform with minimal help from the MLP team and have the confidence that the tools that they use at each stage of the model lifecycle are the correct ones.</li><li><strong>Open Communication</strong>: There’s no point building anything if nobody uses it — to that end, our team has twice-weekly Office Hours where anyone can drop by to troubleshoot more detailed questions that can’t be resolved via messaging. In addition, the team holds biweekly meetings with data science teams to understand their needs and to prevent going off-course. Because we deliver new features every other month, this constant communication is crucial.</li><li><strong>High Adoption</strong>: We are working on integrating our SDK with widely-used Nordstrom tools such as <a href="https://github.com/Nordstrom/nordypy">NordyPy</a>, allowing for a comfortable user experience. As the platform develops, the barrier to entry will get lower and lower. This allows for consumers of the platform to come from across Nordstrom —<em> not just data science and engineering</em>.</li></ul><h3>What’s Next?</h3><p>In future articles, we’ll discuss the future of Cerebro, upcoming features, and the platform’s open-source potential.</p><blockquote>We’d like to thank Nate Lee — another ML Platform engineer who helped write this article — as well as Rizky Ramdhani, Xingtai Li, and Owais Malik, ML Platform team members who provided valuable feedback.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b077b7f212b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[So Your New Team Has Bad Documentation]]></title>
            <link>https://medium.com/@aribray/so-your-new-team-has-bad-documentation-d4c9040ed579?source=rss-2783dcac0d93------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4c9040ed579</guid>
            <category><![CDATA[documentation]]></category>
            <category><![CDATA[technical-writing]]></category>
            <category><![CDATA[diversity]]></category>
            <category><![CDATA[nordstrom]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Ariana Bray]]></dc:creator>
            <pubDate>Wed, 05 Feb 2020 16:31:01 GMT</pubDate>
            <atom:updated>2020-02-05T16:31:01.249Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Learn quickly, contribute early, and don’t get stuck fixing everything yourself.</em></p><figure><img alt="A person writes in a notebook, adding items to a checklist" src="https://cdn-images-1.medium.com/max/1024/0*EOxcoZ3Y62Heg726" /><figcaption>Photo by <a href="https://unsplash.com/@glenncarstenspeters?utm_source=medium&amp;utm_medium=referral">Glenn Carstens-Peters</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>First of all, who am I? I’m an entry-level, Black, female software engineer. I received my degree in Bioethics and Society from the University of Pennsylvania, then made a career switch into tech by attending <a href="https://adadevelopersacademy.org/">Ada Developers Academy</a> in Seattle. Some of my worries or concerns may not resonate with everyone who reads this article, but they are valid experiences that may feel familiar to a lot of people. The point is this: engineers of all types and at every level have a responsibility to contribute to documentation — both team-internal documentation and documentation for the users of their products.</p><p>Picture this: I’ve just started my first engineering job. I have my new machine, my potted plant (Josephine) is at home on my desk, and I’m finally ready to begin working. What better way to start than by walking through my team’s documentation? <strong><em>But… hold up. </em></strong>The documentation provided is confusing and out of date.</p><p>If you’re like me, not being able to work through documentation may unearth some uncomfortable feelings — frustration and our dear friend <a href="https://www.merriam-webster.com/words-at-play/what-is-impostor-syndrome">Imposter Syndrome</a> among them. Maybe you’re worried that bringing up issues early will get you labeled as difficult. If you’re <em>really </em>like me, you aren’t going to let poor documentation get in the way of establishing yourself as an active, productive member of your team.</p><p>So, what do we do about it? We don’t have to suffer through the confusion of poor documentation just because everyone else has struggled. Here are some things I’ve learned that helped me turn the setback of poor documentation into an advantage that I use to contribute early <em>and </em>learn quickly.</p><blockquote><strong>Quick Tip: Hold On To Your Receipts. </strong>A tip that I’ve received from other <a href="https://medium.com/@makemuse/intersectional-feminist-vocab-using-the-term-womxn-b0076aa99794">womxn</a> engineers is to always keep a record of my work. Oftentimes (and for myriad reasons), our accomplishments are overlooked or minimized. While I haven’t experienced this at Nordstrom, I have noticed it at other points in my career. I counteract that by making sure that every one of my work-specific tasks has a story, ticket, or some kind of permanent record.</blockquote><h3>Before You Even Touch The Documentation…</h3><p>Transform your thinking! On my first day, I felt embarrassed and annoyed that I understood basically nothing about my team’s product (Was it a logical feeling? No. Did I feel it? <em>Absolutely</em>). As it turns out, having no context <em>wasn’t a disadvantage after all. </em>I was able to look at the existing documentation with fresh eyes. We all make assumptions. Whoever wrote the documentation made assumptions about a user’s base level of knowledge. I didn’t have this knowledge, so I was able to see areas for clarification and opportunities for an improved user experience.</p><h3><strong>Open Pandora’s Box — but <em>take notes</em>!</strong></h3><p>It’s going to be painful, but the first thing to do (after creating a record of what work is being done) is to work through the existing documentation. My manager gave me this task as part of my onboarding. Keeping a notebook and pencil nearby kept following the documentation from becoming an exercise in frustration. At each step, I made a note of problem areas.</p><p><strong>What’s a problem area?</strong> A problem area can be an area where you (or a user) will second-guess or get stuck on what to do next. A problem area can also be an area with typos, grammar mistakes, or broken links. Is something in the documentation vague? Is the documentation out of date? Make a note of these things and move on. If the documentation is so poor that getting to the next step is impossible, note that too.</p><h3><strong>Don’t Go Too Far Down the Rabbit Hole (also titled Don’t Typecast Yourself)</strong></h3><p>Thorough documentation is rarely confined to one page. It’s tempting to follow every link, writing down things that should be improved. Documentation should absolutely be clarified, but taking on too much at once can be dangerous. If you’re not careful, you could end up spending the majority of your time fixing everyone else’s documentation. I know that I’m actively avoiding being pegged as the “documentation girl.” We are engineers. We’re here to write <strong>code</strong> (with accompanying documentation, of course). Rather than take all the problems on yourself, discuss setting a team standard for documentation. For example, the team could have a rule that states, “Any time a section of code is updated, its accompanying documentation must be updated before a merge request is approved.”</p><h3><strong>Be the Change</strong></h3><p><strong>Bring up the solution while bringing up the issue</strong>. At the next team meeting or standup, speak up about your onboarding or documentation experience. I said something like, “I’ve been getting up to speed by reading the quick start documentation that we provide to our users, and I’ve noticed some difficult areas that can be updated or clarified. I want to take on that task.” That is more helpful than commenting, “This documentation is horrible and confusing, and I couldn’t figure out how to use the service. Please tell me what to do.”</p><p>Once you have ownership of <em>your part</em> of the solution, remember to document that work. If your team uses a Kanban board, create a story or stories that show what you’re working on. For example, “Update new team member onboarding documentation” or “Clean up external FAQs”.</p><h3><strong>It’s Dangerous To Go Alone</strong></h3><p>There’s a reason I didn’t start trying to solve the problem areas as soon as I found them. Maybe the problem areas were issues that I knew how to resolve — things such as grammar or typos. What happened more often, however, was that I needed a second (or third) pair of eyes to look at a problem area. When I nominated myself for some documentation cleanup, I added another sentence — “[Teammate], can I put some time on your calendar this week to walk through the existing documentation?”</p><blockquote><strong>Here’s a hidden benefit</strong>: I got a much deeper understanding of the team’s projects and products because I walked through documentation with a team member. A person who has been on the team for [more time than me] will have context that I’m missing. Instead of waiting weeks or months to acquire that context, <strong><em>boom</em></strong>! I have it now. My teammate also brought up problem areas that I didn’t encounter.</blockquote><p>Because I noted my problem areas before putting time on someone’s calendar, the time I spent with my coworker was much more productive. When I walked through the documentation with my teammate, I brought up those areas: “Step 4 tells the user to do [XYZ], but it looks like they’ll need special permissions. How are those permissions granted? Does the user request them, or do we grant them ourselves?” Write down the solution for each issue. If follow-up with another team member is needed, note that too.</p><h3><strong>Break It Down</strong></h3><p>Bad documentation can seem like an insurmountable problem — where do you even begin fixing it? I approached the documentation the same way I approach most of my problems — by breaking it down into smaller pieces. Let’s say you’re fixing some documentation about how to run an application locally. Within the story you’ve created, create sub-tasks that break the larger issue down into smaller pieces. <em>Remember to be aware of any assumptions you’re making about the reader of the documentation</em>. Are they developers? Is it their first day too? Start at the beginning and create the smallest tasks you can. Here are some examples:</p><ol><li><strong>For team-internal documentation, </strong>if your team prefers using a certain IDE or package manager, make sure that the documentation states that.</li><li><strong>For both internal and external documentation, </strong>if there are dependencies required to make an application or service run, list them (and link to or provide resources on how to install them).</li></ol><h3>Define Done</h3><p>There will be times when poor documentation will take more than one sprint to fix. I’ve adopted a tip from my PM, Oz — I make sure that each story I create (documentation or otherwise) has a <strong>Definition of Done (DOD)</strong>. Having a DOD makes me certain that I’ve ended my documentation cleanup in a logical place. My primary career goal at this moment is to learn and to push code. I’m always picking up tasks on new projects. As much as I believe that fixing documentation is important, I don’t devote 40 hours a week to only that.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/471/1*g1i36o8q2Qzy5pbvBHdIuw.png" /><figcaption>Example story with tasks and Definition of Done (DOD)</figcaption></figure><h3><strong>Get a Proofreader (also titled Two, or Three, or Four Heads Are Better Than One)</strong></h3><p>Whenever I make documentation changes, I open a merge request and require at least one approver. I do this for two reasons.</p><p>Firstly, the team is reminded that documentation changes are coming. Someone else on the team has the chance to review the new documentation. They can catch opportunities for further clarity or spot any issues that I may have missed.</p><p>Secondly, opening a merge request is more confirmation that I was doing valuable work.</p><h3><strong>Share It!</strong></h3><p>Once the documentation has been updated, let people know! For example, you can send an email, highlight it in a meeting, or send a Slack message to your team or to any channels related to your project or product.</p><h3><strong>Be Open to Feedback</strong></h3><p>Once the updated documentation is shared, provide an avenue through which readers and users can suggest feedback. I work with a lot of data scientists. I need to know a fair amount about data science. <strong>I am not a data scientist</strong>. A data scientist might have different problem areas, and that’s okay. My goal is to add value by making documentation clearer — for everyone.</p><blockquote>Thanks for reading along! These tips for handling documentation were compiled from personal experience and tips from friends and coworkers. If there’s anything you do to handle a poor documentation experience, let me know! After all, being open to feedback is an important step.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4c9040ed579" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>