<?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[Plum-living - Medium]]></title>
        <description><![CDATA[Learnings from the team behind Plum - Medium]]></description>
        <link>https://medium.com/plum-living?source=rss----af00eb20b598---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Plum-living - Medium</title>
            <link>https://medium.com/plum-living?source=rss----af00eb20b598---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 14:58:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/plum-living" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Prototyping our VFX-inspired Configurator]]></title>
            <link>https://medium.com/plum-living/prototyping-our-vfx-inspired-configurator-4cd0af00453f?source=rss----af00eb20b598---4</link>
            <guid isPermaLink="false">https://medium.com/p/4cd0af00453f</guid>
            <category><![CDATA[3d]]></category>
            <category><![CDATA[vfx-compositing]]></category>
            <category><![CDATA[vfx]]></category>
            <category><![CDATA[gpu]]></category>
            <dc:creator><![CDATA[Xavier Ayme]]></dc:creator>
            <pubDate>Fri, 05 May 2023 09:55:59 GMT</pubDate>
            <atom:updated>2023-05-09T09:16:26.060Z</atom:updated>
            <content:encoded><![CDATA[<h4>Use of compositing technique for near real-time image generation</h4><p><em>In the first article of this series, I shared the genesis of the </em><a href="https://plum-living.com/l/styler?utm_source=medium&amp;utm_type=article"><em>Plum styler</em></a><em> and how we dismissed existing technologies to create our own 3D inspiration tool. Keep reading to discover our journey to create the Plum styler.</em></p><p>After dismissing the two main technologies available on the market — real-time 3D and precomputed 3D — we had no other option than to create our 3D inspiration tool from scratch. But where to start? Our breakthrough came from a discussion with Julien, a 3D and VFX expert, and who became Plum’s Lead 3D Artist along this journey. Thank you Julien for making all this happen 🙏.</p><h3>Using the VFX Principles of Compositing</h3><p>In the VFX industry, one technique often used to create images is called “<a href="https://www.studiobinder.com/blog/what-is-compositing-definition/">compositing</a>”. To explain it in broad terms, compositing works like a digital collage in which you would take pixel scissors and glue to combine elements and build a single image. The simplest, most-known compositing technique is the use of a green screen to incorporate a background in a video — like in the meteo report on TV, or the last Marvel movie 😊</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M1zqx-z4rpXil7EqyqWI0w.png" /><figcaption>Compositing of a Tap &amp; Kitchen environment</figcaption></figure><p>What if we used this technique to create our inspiration tool?</p><p>In our case, we would need to sum up all the kitchen elements (fronts, handles, taps, etc.) pixel by pixel, <strong>including their respective contributions</strong><em> </em>to other objects (reflection, shadow, color bleeding) to create a photorealistic picture of the selected design. Our tool would then build the resulting image and display it on our website interface.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QqkUUwp2aFX_5LuwLIo1dA.jpeg" /><figcaption>Different levels of contributions explained</figcaption></figure><p>See how changing the tap modifies the environment around it:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D-gBXI_Y_N4XgXCNAa6WdQ.jpeg" /><figcaption>Changing the tap influences the environment around, and the other way</figcaption></figure><p>The concept is enticing, but it requires carefully selecting dependencies between variables to prevent the number of possible combinations from exploding.</p><p>Indeed, in a kitchen, fronts and handles are dependent on each other, but, for example, high cabinets and base cabinets are independent: the cross-shadows and light reflection can be ignored without compromising the photorealism of the image. Therefore, the generated picture will be the sum of the top and bottom parts, each being the combination of the variables (= objects that compose the scene) and their contributions.</p><p>All we have to do is build it 😄.</p><h3>Building our VFX-based Inspiration Tool</h3><p>The compositing technique is a 2-step process:</p><h3>Step1: Generating image parts</h3><p>Based on VFX techniques and software, the “Styler pipeline” takes all of the variables and automatically generates unique image parts. Each image is rendered thanks to a render farm — a data center optimized for 3D rendering — and decomposed into its <em>atomic </em>parts.</p><p>For this stage, we decided to partner with <a href="https://qarnot.com/en">Qarnot Computing</a>, a French company that offers low-carbon and high-performance computing power. Selecting a data center that takes advantage of ​​computer waste heat seemed obvious. And Qarnot Computing has been great with its support (many thanks Ariane!) and was natively integrated with our 3D rendering software.</p><p>Once the rendering is completed, the 100.000 unique image parts are stored in the cloud, ready to be picked by our recomposition service.</p><h3>Step 2: Achieving near Real-Time Compositing</h3><p>Our recomposition service is a C#-based microservice, hosted by Azure. It performs various operations sequentially when it receives a request to generate a configuration:</p><ol><li>it computes the different image parts, ready to be “glued” together,</li><li>downloads and decompresses them,</li><li>sums them up <strong>pixel by pixel, </strong>channel by channel (RGB) to create a resulting bitmap,</li><li>and uploads the result to our CDN (Content Delivery Network), ready to be displayed in the client’s browser</li></ol><p>For instance, our demo kitchen is composed of:</p><ul><li><em>Beige Rosé </em>matt lacquered fronts and natural oak sides for high blocks</li><li><em>Bleu Nuit </em>matt lacquered fronts for low blocks</li><li><em>Mini-rond</em> brass handles</li><li>A brass plinth</li><li>and a <em>Loop</em> brass mixer tap.</li></ul><p>Our recomposition service will sum up all individual images corresponding to the requested configuration (between 50 to 100 on average) and send the result to the customer’s browser.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WPW0m0V_eIMAyQ4sEzzCAA.png" /><figcaption>Overall image is a sum of different independant blocks ‘glued’ together</figcaption></figure><p>Are you still following? 😅</p><h3>Developing our First Prototype</h3><p>After 6 months of development for Julien &amp; myself, a lot of hurdles have been cleared and we finally released our first working prototype, capable of managing 100.000 billion combinations with three different points of view and returning a photorealistic picture✌️.</p><p>But there’s still a minor issue: it takes 45 seconds to compute a single image 😬.</p><p>We started investigating and found a hidden pitfall. Remember: for each configuration, our service sums up about 100 images, each being 100 ko. It seems to be an acceptable amount of data for near real-time processing. So why does it take so long?</p><p>We eventually found that this calculation does not take into account the compression factor: each image is around 100 ko after JPEG compression but the algorithm needs uncompressed byte arrays in input. After decompression, we get about 100 images, each being 1800 x 1350 pixels, x 3 colors per pixel (red, green, blue), x 1 byte per color = 700 Mb to handle instead of 10 Mb!</p><p>Unfortunately, the C# language (and its Bitmap class) is too high level to manage efficiently such low-level parallel operations. Our prototype remains incompatible with real-time usage on a graphic interface 😓.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X7LHZXUL2OG5woZXAWS_xA.jpeg" /><figcaption>Our first Styler implemented ‘demo’ Kitchen</figcaption></figure><p>In the next article, I’ll share the solutions we implemented to drastically speed up the recomposition process.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4cd0af00453f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/plum-living/prototyping-our-vfx-inspired-configurator-4cd0af00453f">Prototyping our VFX-inspired Configurator</a> was originally published in <a href="https://medium.com/plum-living">Plum-living</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Plum Styler Journey]]></title>
            <link>https://medium.com/plum-living/plum-styler-journey-39b5d4fb8c01?source=rss----af00eb20b598---4</link>
            <guid isPermaLink="false">https://medium.com/p/39b5d4fb8c01</guid>
            <category><![CDATA[vfx]]></category>
            <category><![CDATA[3d]]></category>
            <category><![CDATA[vfx-compositing]]></category>
            <category><![CDATA[gpu-computing]]></category>
            <dc:creator><![CDATA[Xavier Ayme]]></dc:creator>
            <pubDate>Fri, 05 May 2023 09:51:45 GMT</pubDate>
            <atom:updated>2023-05-05T10:09:11.732Z</atom:updated>
            <content:encoded><![CDATA[<h4>Birth of a 3D Inspiration Tool at the Service of Design</h4><p>Plum was born in Paris in 2020 to allow everyone to live in a place they have created. Our mission is to <strong>inspire our clients, guide them</strong> <strong>and offer them the best products</strong> to create the interior they dream of. On our website, you can find countless facades, accessories, and paints to personalize an Ikea kitchen or dressing room. But what people don’t always see at a single glance is the cutting-edge tech that we built to better inspire &amp; serve our clients.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3qaXN4Ne11NouNBJVv2Nvw.jpeg" /><figcaption><em>This kitchen does not exist (yet!)</em></figcaption></figure><h3>From a Client Need to a Tech Problem</h3><p>Defining ourselves as an inspirational brand sets the bar high in terms of qualitative visuals. We needed to display high-quality images in line with our brand positioning to help interior designers &amp; clients project themselves in the interior they are designing.</p><p>We wanted our clients to be able to visualize their kitchen, bathroom, or dressing if they combined one front color and shape with a specific handle. But with the number of colors, designs, and finishes Plum offers combined with all accessories, the possible combinations were extremely numerous 🤯, which made it quite difficult for our clients to project themselves.</p><p>So we started wondering,<em> </em>how could technology help our clients better visualize their design? We started dreaming of a 3D-based tool that could display a photorealistic visualization of the client’s inspiration (model, shape, color, and accessories). We could not imagine embedding the inspiration into their own environment, this was out of our technical-reach at the time, but at least into various preset layouts close to actual customer projects.</p><p>But before reinventing the wheel, let’s try to see what was available on the market.</p><h3>Assessing Two Main Technologies</h3><p>We quickly identified two possibilities: real-time 3D and precomputed 3D. Can either of them tackle our needs?</p><h3>Real-time 3D for web browser</h3><p>Real-time 3D refers to numerous libraries and techniques — you may have heard of WebGL, babylon.js, or three.js for example. In practice, the client’s browser handles the 3D digital model, and its rendering is performed by the computer that “displays the 3D”.</p><p><em>Note: If you are not familiar with the term “rendering”, the idea is to create a picture similar to what the human eye or a camera sees from the interaction between a 3D model (the real world) and billions of light rays emitted by light sources — like the sun or a lamp. Their reflection and diffusion on the 3D model and its surroundings are computed. The process is called “rendering”, and the output is a 2D image.</em></p><p>Real-time 3D offers total flexibility on the 3D model and the number of light sources interacting with it. However, to be fluid for the customers, the rendering would need to happen in less than a fraction of a second, otherwise, the user would feel a “lag”. But even with the best laptop available on the market, it limits the degree of photorealism that can be achieved: the computer basically does not have enough time and computation power to emulate enough light rays and reflections. This is a trade-off gamers are well aware of: remember how video games 🕹looked like 10 years ago? And 20 years ago?</p><p>Said in a simple way, and leaving aside technical progress made in this science, photorealism is a function of two simple parts: the computation power available (a.k.a. the computer speed) and the time it takes for rendering.</p><p>Unfortunately, in 2023, despite the huge progress made every year by the manufacturers (cf. <a href="https://en.wikipedia.org/wiki/Moore%27s_law">Moore’s Law</a>), an average computer cannot achieve photorealism in near real-time. It may be the case in a couple of years, but it is not yet possible.</p><p>This was a dealbreaker for us. The other option we had in hand was precomputed 3D.</p><h3>Precomputed 3D</h3><p>With precomputed 3D, all images are precomputed on a rendering server and stored in the cloud. Coming back to my previous assumption of the computation power and the rendering time necessary to achieve photorealism:</p><ul><li>we can use dedicated made-for-purpose servers (so computation power is increased significantly)</li><li>and we have plenty of time! (let’s say 24 hours, compared to a fraction of a second)</li></ul><p>Under those favorable conditions, we can definitely achieve photorealism and render a picture that cannot be differentiated from a photograph.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*dlAcyw2EDUCuzxaX2vb4aA.png" /><figcaption><em>This plant is a precomputed 3D image of a plant. Would you have known it?</em></figcaption></figure><p>How could we implement it? Whenever a client selects a configuration, the website calls the corresponding picture and displays it on the client’s screen. Easy life, quite a simple web service to develop!</p><p>But in our simple <a href="https://plum-living.com/l/styler/app?id=kitchen_1&amp;utm_source=medium&amp;utm_type=article">demo kitchen</a>, for a single front, with 30 colors available, 4 designs and 5 finishes, a split between high and low blocks, and a selection of accessories like handles, taps, plinths, this gives us about … 100.000 billion possible combinations and therefore 100.000 images to generate and store.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YXiNDfK49SEXe2dSHxBmMw.jpeg" /><figcaption>Our demo Kitchen and its possible variations</figcaption></figure><p>Is it feasible? Obviously not:</p><ul><li>Each image has an average size of ~ 100 ko</li><li>100 ko x 100.000 billion images leads to 10.000 billions of Mo, so <strong>10 Exabytes of data</strong>,</li><li>So basically, to store all the combination images from our <em>simple</em> kitchen configuration, we would need… a whole Datacenter and a few hundred million euros to render them.</li></ul><p>This goes far beyond Plum’s purchasing power 💰😝.</p><p>In this series of articles, I’ll share our journey to create the <a href="https://plum-living.com/l/styler?utm_source=medium&amp;utm_type=article">Plum style</a>r — behind the scene.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=39b5d4fb8c01" width="1" height="1" alt=""><hr><p><a href="https://medium.com/plum-living/plum-styler-journey-39b5d4fb8c01">Plum Styler Journey</a> was originally published in <a href="https://medium.com/plum-living">Plum-living</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a semantic layer in Preset (Superset) with dbt]]></title>
            <link>https://medium.com/plum-living/building-a-semantic-layer-in-preset-superset-with-dbt-71ee3238fc20?source=rss----af00eb20b598---4</link>
            <guid isPermaLink="false">https://medium.com/p/71ee3238fc20</guid>
            <category><![CDATA[superset]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[data-vizualisation]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[dbt]]></category>
            <dc:creator><![CDATA[Pierre Vanacker]]></dc:creator>
            <pubDate>Wed, 15 Feb 2023 11:06:27 GMT</pubDate>
            <atom:updated>2023-02-15T11:06:27.291Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v8lia2acSEV47hrZH-H5Tg.jpeg" /></figure><p>Plum is an interior design &amp; decoration start-up. At Plum, we unleash our customers’ creative potential, by providing them with meaningful and inspirational DIY and interior design content, top-of-the-line products, and expert advice and guidance. Ultimately, we want to be the creative partner of choice when it comes to interior design.</p><p>For that, our team needs to perfectly know our consumers &amp; community behaviors, our suppliers data… it means synchronize data from different tools, and define the data &amp; metrics we want to expose to each team.</p><p>To kick off our Data Platform creation process, we had to benchmark the Data Visualization solutions. <strong>Preset (a fully-managed service for Apache Superset) was a good match, with a balance between simplicity and versatility.</strong></p><p>Amongst all the features offered by <strong>Preset, the synchronization feature with dbt grabbed our attention.</strong></p><p>Another reason we favored Preset was its open-source tool, which avoided locking Plum into a proprietary system.<strong> The synchronization with dbt would take us one-step further</strong>, allowing us to define our entire exposed data model not only by code but also with a solution that was totally independent of the visualization platform.</p><p>This article will show how we implemented this project technically, and the advantages it provided Plum.</p><h3>What is a metric definition ?</h3><p>A <em>metric</em> is defined by dbt as “an aggregation over a <a href="https://docs.getdbt.com/terms/table">table</a> that supports zero or more dimensions”. More generally, it’s a representation of one indicator over a dataset, linked to one or many company KPIs.</p><p>Creating metrics in the modelling layer is useful as it gives us a single definition of an indicator understandable across business teams. We could say it’s a first step of building a semantic data layer.</p><p>For example we could provide a metric which computes the revenue, taking into account all necessary fields :</p><pre>  - name: total_revenue_vat_included<br>    label: Total revenue VAT included<br>    model: ref(&#39;m_orders&#39;)<br>    description: &quot;The total revenue (VAT included). It filters out cancelled and closed orders&quot;<br><br>    calculation_method: sum<br>    expression: total_after_discount_vat_included<br><br>    timestamp: transaction_date<br>    time_grains: [day, week, month, quarter, year, all_time]<br><br>    filters:<br>      - field: status<br>        operator: &#39;!=&#39;<br>        value: &quot;&#39;cancelled&#39;&quot;</pre><h3>How it works</h3><p>Here are the two main functionalities of this system:</p><ul><li>Whenever a developer pushes a new metric definition in dbt, Preset is updated with this metric definition</li><li>Whenever a developer updates a table schema, Preset dataset definitions are updated</li></ul><p>This allows us to define metrics &amp; schemas by code and always maintain Preset in sync with the dbt configuration. It also allows us to expose only selected datasets to the users, while keeping them in the freshest state, and hide raw &amp; cumbersome utility tables under-the-hood.</p><p>We use a dbt-core project for this, hosted on a Github repository.</p><p>Here is the architecture structure of the synchronisation from Github to Preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tKBLRN_15FdMbAEmFxg3vA.png" /></figure><p>We rely on Github actions to push the updated metrics &amp; schema on the appropriate workspace every time a developer pushes something on specific branches.</p><p>Basically:</p><ul><li>pushing on the development branch updates the <em>development</em> workspace of our Preset account</li><li>pushing on the main branch updates the <em>production</em> workspaces of our Preset account.</li></ul><p>This schema sums up the architecture behind Preset workspaces:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_lrUsHmn3k7kyOwJxRgKoA.png" /></figure><h3>Focus: CI configuration</h3><p>The CI configuration has been greatly inspired by this article from Preset: <a href="https://preset.io/blog/dbt-superset-integration-preset/">https://preset.io/blog/dbt-superset-integration-preset/</a></p><p>It relies on the Preset CLI (which works over Superset API) which provides the synchronization utility: <a href="https://github.com/preset-io/backend-sdk">https://github.com/preset-io/backend-sdk</a></p><p>The necessary steps:</p><ul><li>Retrieve the Preset CLI repository (you can also host it internally and/or create a package) + Install dbt</li><li>Execute a <em>dbt compile </em>command<em> </em>to generate the manifest.json file. Preset CLI indeed has to parse the dbt manifest to retrieve the structure of your models.</li><li>Publish the resources you need with the CLI!</li></ul><p>The idea is to select only the models/tables you want to expose to your users (put raw tables aside for example &amp; only show datamarts). Metrics defined in dbt will automatically be accessible in Preset as metrics linked to your tables.</p><p>Here is a <em>Github Actions</em> file example for this process:</p><pre>name: &#39;Publish dbt resources to Preset&#39;<br><br>env:<br>  DBT_GOOGLE_PROJECT: xx<br><br>on:<br>  push:<br>    branches:<br>    - dev<br>    - main<br><br>jobs:<br>  publish:<br>    name: &#39;Publish resources&#39;<br>    runs-on: ubuntu-latest<br><br>    defaults:<br>      run:<br>        shell: bash<br><br>    steps:<br>    - name: &#39;Checkout CLI&#39;<br>      uses: actions/checkout@v3<br>      with:<br>        repository: preset-io/backend-sdk<br>        path: preset-cli<br><br>    - name: &#39;Install CLI&#39;<br>      run: |<br>        cd preset-cli &amp;&amp; pip install . &amp;&amp; cd ..<br><br>    - name: &#39;Checkout&#39;<br>      uses: actions/checkout@v3<br><br>    - name: &#39;Install dbt&#39;<br>      run: |<br>        pip install dbt-core==1.3.1 dbt-bigquery==1.3.0<br><br>    - name: &#39;Define variables based on branch&#39;<br>      run: |<br>        if [[ $GITHUB_REF == &#39;refs/heads/main&#39; ]]; then<br>            echo &quot;WORKSPACE=yy,zz&quot; &gt;&gt; &quot;$GITHUB_ENV&quot;<br>            echo &quot;DBT_TARGET=prod&quot; &gt;&gt; &quot;$GITHUB_ENV&quot;<br>        else<br>            echo &quot;WORKSPACE=zz&quot; &gt;&gt; &quot;$GITHUB_ENV&quot;<br>            echo &quot;DBT_TARGET=dev&quot; &gt;&gt; &quot;$GITHUB_ENV&quot;<br>        fi<br><br>    - name: &#39;Compile manifest&#39;<br>      run: |<br>         dbt deps &amp;&amp; compile --target=${DBT_TARGET}<br>      env:<br>        GOOGLE_APPLICATION_CREDENTIALS: ${{ secrets.DBT_GOOGLE_BIGQUERY_KEYPATH }}<br>      working-directory: dbt-models/plum_core_data<br><br>    - name: &#39;Publish refined resources&#39;<br>      env:<br>        PRESET_API_TOKEN: ${{ secrets.PRESET_API_TOKEN }}<br>        PRESET_API_SECRET: ${{ secrets.PRESET_API_SECRET }}<br>      run: |<br>        preset-cli --workspaces=${WORKSPACE} superset sync dbt \<br>          ./target/manifest.json \<br>          --profiles=./profiles-ci.yml \<br>          --project=plum_core_data --target=${DBT_TARGET} --import-db -s tag:refined tag:datamart -x tag:ephemeral<br>      if: github.event_name != &#39;pull_request&#39;<br>      working-directory: dbt-models/plum_core_data</pre><p>Note the synchronization tool is also able to retrieve dashboards created in Preset as dbt <em>exposures.</em> <br>Implementing this<em> (not necessarily in the CI for this part)</em> would greatly help to backup dashboards and/or exposing them to a Data Catalog for example.</p><h3>Conclusion</h3><p>This Preset feature allows us to define the data &amp; metrics we want to expose to our users “as code” with dbt, without being tied to a closed solution or language specific to the visualization tool.</p><p>Something is still missing though: how to collaborate with business users in the definition of the metrics. Editing configuration files in a Github repository is not the best for end users. I’m quite confident this part will be improved very soon.</p><p>In a future article, I’ll detail a bit the architecture of the Data Platform we started to create 4 months ago. Stay tuned!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71ee3238fc20" width="1" height="1" alt=""><hr><p><a href="https://medium.com/plum-living/building-a-semantic-layer-in-preset-superset-with-dbt-71ee3238fc20">Building a semantic layer in Preset (Superset) with dbt</a> was originally published in <a href="https://medium.com/plum-living">Plum-living</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>