<?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 Régis Amichia on Medium]]></title>
        <description><![CDATA[Stories by Régis Amichia on Medium]]></description>
        <link>https://medium.com/@regis_90991?source=rss-508f75e0bd6e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*OULCpxNrNVDnmMlqenwyjA.jpeg</url>
            <title>Stories by Régis Amichia on Medium</title>
            <link>https://medium.com/@regis_90991?source=rss-508f75e0bd6e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 23:21:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@regis_90991/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[How we set up an ML OPS platform from scratch : an example from product categorization.]]></title>
            <link>https://medium.com/@regis_90991/how-we-set-up-an-ml-ops-platform-from-scratch-an-example-from-product-categorization-d8a5f76845a4?source=rss-508f75e0bd6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/d8a5f76845a4</guid>
            <category><![CDATA[categorization]]></category>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[airflow]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Régis Amichia]]></dc:creator>
            <pubDate>Thu, 23 Jul 2020 16:04:12 GMT</pubDate>
            <atom:updated>2020-07-28T16:08:27.043Z</atom:updated>
            <cc:license>http://creativecommons.org/publicdomain/zero/1.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CRJX-DtWaj8G8xO0-0pnXg.png" /></figure><h3>How we set up an ML OPS platform from scratch : an example from product categorization.</h3><p><a href="https://www.foxintelligence.fr/">FoxIntelligence</a> aims to become the leading consumer-centric intelligence company for digital transactions. Thanks to our proprietary panel, we collect data from hundreds of millions transaction emails to provide unseen insights on ecommerce. A transaction email contains the name of the product and its description but never its category. However we need the latter, to be able to rank each product with similar ones, and thus provide <a href="https://www.foxwatch.io/">statistics breakdown by category to our customers</a>. That’s why among the necessary data enrichment, one of our biggest challenge is product categorization.</p><p>The most accurate way to categorize a product is to do it manually. Because of all the subtleties of product taxonomy, we assume humans would always perform better than algorithms. Furthermore, it would only takes few seconds to categorize a product through a web app so let’s go!! No? Well unfortunately if you multiply few by hundreds of millions and soon billions of products, manual categorization would take…</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fgiphy.com%2Fembed%2FxT5LMVhBajuh8pQre0%2Ftwitter%2Fiframe&amp;display_name=Giphy&amp;url=https%3A%2F%2Fgiphy.com%2Fgifs%2Fseason-16-the-simpsons-16x16-xT5LMVhBajuh8pQre0&amp;image=https%3A%2F%2Fmedia3.giphy.com%2Fmedia%2FxT5LMVhBajuh8pQre0%2Fgiphy.gif&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=giphy" width="435" height="328" frameborder="0" scrolling="no"><a href="https://medium.com/media/ebf310de8a2e51a7d82528e444b7945a/href">https://medium.com/media/ebf310de8a2e51a7d82528e444b7945a/href</a></iframe><blockquote><em>Want to read this story later? Save it in </em><a href="https://usejournal.com/?utm_source=medium.com&amp;utm_medium=blog&amp;utm_campaign=noteworthy&amp;utm_content=eid7"><em>Journal</em></a><em>.</em></blockquote><p>We have therefore developed several automatic product categorization solutions including one based on machine learning. After a first version, we understood why deploying machine learning solutions <a href="https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf">involves much more than just slapping a dataset on a Tensorflow model, launching a training and putting the model inside a web server.</a> Monitoring the accuracy of models in production is a crucial aspect of the project to 1) make every stakeholders aware of data quality 2) allow data scientists to spot areas of improvement and iterate quickly on new model versions to increase accuracy. To do so, data scientists need to focus on areas where they add the most value, be less dependent on software engineers to deploy models and efficiently manage every step of the data science life cycle, from data labeling to production monitoring.</p><h3><strong>Data Labeling</strong></h3><p>Our product categorization challenge is a classification task which, as every supervised machine learning project, starts by labeling data. We decided to develop our own categorization taxonomy to encompass every type of product in our panel and to harmonize categories between brands and merchants. Our categorization tree has three levels and more than 2000 leaves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/813/1*5FFUeJhfhoNz6XkkHti0SA.gif" /><figcaption>Foxintelligence french taxonomy</figcaption></figure><p>Because our dataset of products comes from dozens of different catalogs, each one with its own taxonomy, it was necessary to match every categories with our tree. It had to be a manual process because the accuracy of labelled data is a prerequisite for the success of our model, and we couldn’t risk even minor errors at this step. It represents thousands of categories to match, which is very time-consuming so we outsourced it to a manual data processing team in Madagascar which uses a web app we developed internally.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PYxqogZ9cCO1-Yj48sPabQ.gif" /><figcaption>web application used by the data processing team to label data</figcaption></figure><p>This method allow us to quickly build a dataset of millions of products in different languages. For each language, we will build a specific model and we aim to reduce the time of a complete learning run as much as possible.</p><h3><strong>ML Orchestration</strong></h3><p>Most tools built for machine learning tend to focus on the “fun parts” of machine learning (training, deployment etc.). As this diagram shows, the scope of machine learning systems is much larger.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/974/1*vsALP2uaWHM5sM8pqeurDw.png" /><figcaption><a href="https://cdn.confluent.io/wp-content/uploads/hidden-tech-debt-ml-system.png">https://cdn.confluent.io/wp-content/uploads/hidden-tech-debt-ml-system.png</a></figcaption></figure><p>We needed to solve three problems with our new platform for orchestration:</p><ul><li>The inability to iterate quickly. We wanted to make sure Data Scientists spent as little time as possible battling incidental complexity.</li><li>Insufficient ownership of the complete flow, relying too much on other teams every step of the way.</li><li>Performance and quality issues were rarely identified early on. Manual checks during the training pipeline can only get you so far. These problems are exacerbated when data scientists aren’t responsible for pushing their models to production. Engineers without domain specific knowledge cannot spot problems as well as data scientists.</li></ul><h4>Why Airflow + Kubernetes</h4><p>Our pipelines need to be idempotent, reproducible and observable. After several benchmarks of open source and SaaS solutions we landed on <a href="https://airflow.apache.org/">Airflow</a><em> </em>for the following reasons:</p><ul><li>An extensive web UI which allows Data Scientists to trigger and monitor pipelines easily.</li><li>An API so it can be integrated into existing systems.</li><li>Official support for Kubernetes as an executor and operator to run tasks inside docker images, reducing dependencies between dags and teams. Furthermore, Airflow’s base cost when running on k8s is close to zero as Airflow’s workers can scale to zero when no dags are running.</li></ul><p>We initially decided to distribute the code for our pipelines via private PyPI packages which would then be used by Airflow. As projects started to add up, we slowly sank in dependency hell and it became unmanageable. What happens when a Dag needs Tensorflow 2.0 and another Tensorflow 1.10, yet they are all deployed within the same docker image ?</p><p>This is why we rely extensively on Airflow’s Kubernetes Operator to reduce friction between dags. Dags are packaged in one or several docker images and Airflow only acts as a scheduling and monitoring tool. <a href="https://medium.com/bluecore-engineering/were-all-using-airflow-wrong-and-how-to-fix-it-a56f14cb0753">This article sums up well the advantages of running Airflow with the Kubernetes Operator.</a></p><p>When a data scientists makes changes to their code, Jenkins automatically picks them up in GitHub and packages the pipeline as a docker image. They can experiment quickly with different branches on GitHub which get their own docker tag and are available in airflow. Data scientists can execute multiple experiments at the same time without impacting other teams as dags are isolated.</p><p>We run Airflow on EKS, the managed Kubernetes solution by AWS. Data scientists can request special machines like GPU instances if their workload requires it. We use spot instances, since training takes long hours we don’t care if the initial time to get an instance is high. It reduces cost significantly.</p><p>Now that we have built our orchestration system, let’s talk about what’s in it for our product categorization case.</p><h3><strong>Learning Pipeline</strong></h3><p>The training process is divided in multiple stages to have simple code blocks we can easily maintain and evolve. Each step is independent, so if it crashes it does not have to start over the whole pipeline. It also means that multiple pipelines can run alongside each other. Each step is an ETL process where transformation refers to the usual machine learning steps : data cleaning, text preprocessing, features engineering, model training, model evaluation and validation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wL5VdFblfMw6FnnN6oizMw.png" /></figure><p>The <strong>Ingest</strong> block connects to our internal data sources (BigQuery, S3, MongoDb and SQL) to load data (product names and categories).</p><p>The second step, <strong>data validation</strong>, matches each breadcrumb with our own taxonomy and removes products which can’t be matched because either we do not provide intelligence for those industries (such as weapons) or because there wasn’t enough information on the original website. We also monitor the distribution of categories to make sure it is close enough to the distribution of categories in our panel.</p><p>The <strong>text preprocessing</strong> contains classic NLP processing (lower casing, stop word removal…) but also some more custom transformation, which proved to be effective for our use-case. Now that we have a clean and labelled dataset, let’s build the model.</p><p>The <strong>feature engineering</strong> step aims to create features from our dataset which will be used to train our model. It starts by tokenizing text on a character level. When building a natural language processing model, it is a good idea to choose the smallest unit that our model will understand and that is often words. However, e-commerce product names and description are not proper sentences but often abbreviations. By choosing character level tokenization, our model will be stronger on spelling errors (which can happen on C2C platforms like Vinted or Ebay), abbreviations and rare words. It also gives our model a smaller and less sparse parameter space which makes it converge faster. Using word level tokenization would allow us to capture the semantic meaning, and we could start with pre-trained vectors instead of building our embedding representation from scratch. Building a machine learning model is all about choices and trade-offs. Ultimately character level embedding proved to be the optimal solution.</p><p>We are now ready to <strong>train</strong> our model followed by the <strong>model evaluation</strong> where we compute the confusion matrix and precision/recall/f1 score on the test set at each level of our categorization tree.</p><p>Finally, we compare the evaluation scores of our model with the best one so far in the <strong>model validation </strong>step and if the score is better or at least close enough, the model is deployed on our inference API.</p><h3>Serving</h3><p>Machine learning is a relatively new field compared to traditional software engineering, it hasn’t reached the same level of maturity in many aspects. Our machine learning models have a direct impact on the product we sell our client. We build them with the same requirements we develop other, more traditional services. This includes, for instance, testing, peer-reviewed code and versioning which we will focus on here.</p><p>When a new model makes it through our training pipeline and has passed automated checks, data scientists can assign it a version. As a rule of thumb we consider adding more data to the pipeline as a patch, changing the code of the pipeline as a minor change and the actual structure of the model as a major change.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PoEiUEsQvl2OXgsLB8tphA.png" /><figcaption>first versions of our model</figcaption></figure><p>This system increases visibility on the changes we bring our models. Rollouts of new models used to be all or nothing, we now have the ability to track changes over time, test newer models in production alongside current and past ones and rollback changes if needed.</p><p>Assigning a version number and pushing the model to production can be done with the click of a button. Data scientists do not rely on other engineers who may not have the full context of new models. This process increases the autonomy of data scientists and reduces friction when bringing models to production.</p><p>Our serving API uses AWS Sagemaker so we can deploy and update models via code. As such most of our pipeline from start to finish is versioned and available for engineers to update / fix as requirements evolve. Sagemaker also supports horizontal scaling to easily scale up and down instances. Auto scaling reduces cost (comparing with our previous system where several instances were always up) and allows our serving API to absorb traffic peaks seamlessly.</p><h3>Monitoring</h3><p>Any system no matter how well-designed they are will fail at some point. ML systems even more so as the data coming in changes over time.</p><p>We decided to build our own application to increase visibility on the predictions our models are creating. ML systems are usually regarded as one huge black box especially by other teams whose who rely on model predictions so making them more visual and concrete can help bridge that gap. We used to communicate accuracy scores in a static way across merchants and brands. Not only were we unable to track model performance over time, but model accuracy can vary from merchant to merchant (mainly depending on the wording of each merchant). This would cause our accuracy score to be inaccurate for some industries, which was frustrating for internal stakeholders.</p><h4>Check Model Prediction</h4><p>To increase the granularity of our monitoring, <strong>we needed to bring humans in the loop</strong>. We still assume humans are better than algorithms to categorize products and that’s where our manual data processing team comes in. A subset of products predicted is stored in our back-office to be continuously checked, and corrected if necessary.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5_JtlbKmpHu6iSL27rKaJA.gif" /><figcaption>monitoring tool used by the data processing team</figcaption></figure><p>This way, we can monitor if the model is accurate over all merchants and consistent over time. When we have doubts on certain merchants or categories we can focus our labeling efforts on those categories to assess the models’ true performance in those areas.</p><p>As mentioned briefly before, we also use these data points to train future models. Most predictions that are corrected sit on the decision boundary of a model, so they have a big impact.</p><h4>Accuracy Dashboard</h4><p>We built accuracy dashboards for all models and countries that automatically get updated as predictions come in and they are checked by the data processing team. Once a prediction is checked for one version of a model it will be sent to our inference API for back testing across all model versions. That way we can spot discrepancy between versions easier and our data processing team never checks the same product twice.</p><p>We can monitor for accuracy changes in categories and versions over time but also have an overview of the model performance. Data scientists can leverage that information to focus on the weak spots of a model. Data analysts have a clearer sense of the quality of the enriched data on which their analysis are based.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1022/1*JI6aOo_TL9AgOXC89ft07A@2x.png" /><figcaption>Our Machine Learning Projects Lifecycle</figcaption></figure><h3>Conclusion</h3><p>The product categorization is still an ongoing process but now we are ready to take the next step: deploying models in more languages to provide enriched data in more countries and support FoxIntelligence’s international expansion. The feedback received from other teams on the back-office fuelled our roadmap, and we are going to deploy new features like providing recall/f1-score alongside accuracy.</p><p>Building this platform has not always been an easy journey and it took us 6 months but it was necessary to move into the era of industrialized machine learning and increase trust on machine learning projects.</p><p>📝 Save this story in <a href="https://usejournal.com/?utm_source=medium.com&amp;utm_medium=noteworthy_blog&amp;utm_campaign=tech&amp;utm_content=guest_post_read_later_text">Journal</a>.</p><p>👩‍💻 Wake up every Sunday morning to the week’s most noteworthy stories in Tech waiting in your inbox. <a href="https://usejournal.com/newsletter/noteworthy-in-tech/?utm_source=medium.com&amp;utm_medium=noteworthy_blog&amp;utm_campaign=tech&amp;utm_content=guest_post_text">Read the Noteworthy in Tech newsletter</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d8a5f76845a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deploying a Deep Learning pipeline with Docker on AWS]]></title>
            <link>https://medium.com/@regis_90991/deploying-a-deep-learning-pipeline-with-docker-on-aws-a8513b3326ff?source=rss-508f75e0bd6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/a8513b3326ff</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[docker]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Régis Amichia]]></dc:creator>
            <pubDate>Thu, 13 Sep 2018 13:48:07 GMT</pubDate>
            <atom:updated>2019-04-26T14:27:37.132Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hEbJvltnslRrdEzjWQ7Img.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*f6hLntjZ56xnZ3seCJQ8Zw.jpeg" /><figcaption>Our model was built with Keras and Tensorflow as backend</figcaption></figure><p><em>Note: this article was written with </em><a href="https://medium.com/u/bfbad53e5180"><em>Mohamed Labouardy</em></a></p><p>Deep learning models are tricky to design and tune. Every data scientist can relate to the feeling of relief when discovering that their models perform as well on new data as during the training, test and validation phases. However, this is not an achievement but the beginning of a new turbulence zone: how can a model (which in our case will be in production as a REST API) scale to handle an increasing amount of queries? You might have thousands or even millions of requests simultaneously and this is where the trouble begins.</p><p><strong><em>Note</em></strong><em>: Our architecture is based on the</em><a href="https://www.pyimagesearch.com/2018/01/29/scalable-keras-deep-learning-rest-api/"><em> amazing article of Adrian Rosebrock</em><strong><em> </em></strong></a><em>quoted by François Chollet (the creator of Keras) as</em><a href="https://twitter.com/fchollet/status/958069169588617216"><em> the reference architecture to build a scalable deep learning model-serving API</em></a><em>. As mentioned in the article, we worked with Keras (TensorFlow back-end) Flask and Redis but we wrapped it into Docker containers to scale out/in services based on traffic and AWS Lambda to trigger the deployment process from Jenkins.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/999/1*JFm8diyNtZl1h3M8s-z2Bw.png" /><figcaption>The whole infrastructure</figcaption></figure><p>We decided to build an infrastructure where each part of the process is independent, and can communicate with other parts through a Push and Pull model (AWS SNS and Redis). It is inspired by the microservice architecture, well known in software engineering.</p><p>Our Infrastructure is divided into five parts:</p><p><strong>The learning containers : </strong>the source code of our algorithm is wrapped into Docker containers. It allows a better portability, since our data team members are not working with the same version of libraries. The code can be run on every computer, which was helpful during the training process. Once the model is ready to go into production (when it reaches the minimum accuracy level required), a service will load all the necessary data (in our case it is some <em>.pickle, .h5 .csv and .xlsx</em> object) into <strong>Redis </strong>and <strong>S3 </strong>databases. We definitely intend to improve our model: for every new version, the service will erase the previous objects in <strong>AWS </strong>and <strong>Redis </strong>and replace them by the new ones. We strongly recommend to have several environment deployment (3 in our cases) so each new version can go through a test process before being in production.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/ba6caf737023a91380e4737b7d369b13/href">https://medium.com/media/ba6caf737023a91380e4737b7d369b13/href</a></iframe><p><strong>The application builder:</strong> once new objects are loaded, a message is published through AWS SNS and an AWS Lambda function is triggered. It launches the final unit and integration test and eventually deploys our application through <a href="https://jenkins.io/">Jenkins</a>, an open source automation server.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U5BIOuHmtJnPRHte0RygbA.png" /><figcaption>Lambda on AWS triggered by SNS</figcaption></figure><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b2d563d18e881d76d67e7b858ba4e1d6/href">https://medium.com/media/b2d563d18e881d76d67e7b858ba4e1d6/href</a></iframe><p><strong>The producer containers: </strong>they hold our flask application. The producer has one task: to manage the communication between the client and the consumer. Doing so, it handles the request from a client (in our case it is an other microservice of our infrastructure), sends it to a queue (we use Redis), waits for the request to be handled by the consumer and then sends back the prediction to the client. The producer will stay “on hold” until it receives the answer from the consumer, so we make sure a request is only processed once.</p><p><strong>The consumer containers:</strong> this is where the magic happens. Each consumer container constantly loops on the queue to look for new data to predict by our deep learning model and sends back the response to an other queue. In our case, prediction can take a few seconds (while it only takes a few milliseconds for the producer to send a message to the Queue). It is important to constantly monitor the the activity of the service, because you should scale the consumers and producers independently. Otherwise, you will not optimize the performance of your architecture.</p><p><strong>The monitoring: </strong>In addition to the amount of requests handled by our API, we have to monitor many metrics within our application to quickly detect unusual values, bugs or any other anomaly. We mainly use ELK stack to build interactive and dynamic dashboards and Grafana to monitor the infrastructure resource usage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MHpIqKEwhEq9Vm0jXa43fQ.png" /><figcaption>monitoring the amount of requests</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/913/1*kCLOpBJJzYXgemUmAZbcyw.png" /><figcaption>monitoring resource usage</figcaption></figure><p>There are many other architectures, frameworks and tools to deploy deep learning pipeline, depending of your needs. Like many other startups, we faced the scalability issue and found a way to solve it from scratch.</p><h3>About Foxintelligence</h3><p><em>Foxintelligence delivers the best insights on the latest European ecommerce trends, by unlocking intelligence from hundreds of merchants and thousands of brands e-receipts thanks to a panel of millions of online shoppers. We transform the panel industry as a service, at the service of business leaders and marketing professionals who want to drive growth with the most accurate and freshest competitive and consumer information.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8513b3326ff" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Foxintelligence improves data quality collaborating with Side.co]]></title>
            <link>https://medium.com/@regis_90991/how-foxintelligence-improves-data-quality-collaborating-with-side-co-b4f024c28600?source=rss-508f75e0bd6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/b4f024c28600</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[human-resources]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Régis Amichia]]></dc:creator>
            <pubDate>Tue, 05 Jun 2018 16:50:52 GMT</pubDate>
            <atom:updated>2019-04-26T14:27:49.570Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>How Foxintelligence improves data quality collaborating with Side.co</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9r3URh4FUSNM8uQ3dzwYIA.jpeg" /></figure><p><em>Disclaimer : Hi guys, just to make things clear, the post below is not an advertisement. We just wanted to write about Side because it makes a great difference for us. By the way if a guy from Side reads it, congrats guys you’re doing a really good job.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/240/1*bh1sYzB6XlOctJ9OPFaOMg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/225/1*-IEed6GFM8l2OTKv0OSVyg.jpeg" /></figure><p>Foxintelligence is a market intelligence company. We measure e-commerce activity and use aggregated and anonymized data to provide our clients with the best insights to drive their business. We handle millions of unstructured data points every day and one of the objectives of the data team is to create tools and micro-services to structure them. This is why, together with the dev team we have developed a bunch of tools (data pipelines, machine learning algorithm, monitoring services, alerting system etc…) to make sure we have the best data quality.</p><p>Thank to the amazing skills and energy of our technical team,we manage to get a high level of accuracy but we wanted to go further. Besides, we also need data to feed our machine learning algorithm. Therefore, we could not achieve our objectives without human resources to help us and given the specification of our activity, we need flexible resources, able to speak French and English, and we also need them to understand our business, our issues and sometimes, to take a step back and understand all the complexity of the situation. We also need them to give us a feedback on how to improve some of our tools. At this stage, it is important to understand how crucial it is for a startup in the market intelligence industry to have accurate data and this is why this issue was a thorn in the flesh for our company.</p><p>Side brought the solutions to our issues. The make available dozens of qualified students, we called them “siders”, most of which are in some of the prestigious business and engineering schools or universities, able to work remotely or to come to our office. We can choose the length of the mission and the schedule. Everything is totally customized, and their platform allows us to communicate easily with the siders. Because we are a fast-growing startup, we make decisions quickly and we need resources almost immediately. It is Monday 5pm and we need 4 students in our office tomorrow at 9am? But we do not have enough laptops for them? Not a problem at all. Side provides qualified resources very quickly and they bring their own materials.</p><p>We have done more than 20 missions with Side and worked with more than 30 different siders. We know their strengths and points of improvement, so we can choose the siders we want to work with depending on how they perform in their previous missions or we can look at the comments made by other companies where the siders already worked.</p><p>The administrative part does not take any time because in fact, we have almost nothing to do. The siders work as freelance and, once the mission is done, they invoice us through the platform of Side. We registered our credit card so it does not take more than 30 seconds to pay them. And because the price of the mission is known upfront, there can’t be any surprise. Anyway, every time we had a question about the process, their customer service was very quick to answer.</p><p>From our point of view, Side is the best solution today to collaborate with temporary workers for any kind of short mission you can imagine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b4f024c28600" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>