<?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 David de Meij on Medium]]></title>
        <description><![CDATA[Stories by David de Meij on Medium]]></description>
        <link>https://medium.com/@daviddemeij?source=rss-d9c8536f89ac------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*4RpfZ53xtWxjxIaxdGgKqA.png</url>
            <title>Stories by David de Meij on Medium</title>
            <link>https://medium.com/@daviddemeij?source=rss-d9c8536f89ac------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 23:21:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@daviddemeij/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[Engineering Biology: The Quiet AI Revolution in Protein Design]]></title>
            <link>https://medium.com/@daviddemeij/engineering-biology-the-quiet-ai-revolution-in-protein-design-b67d587a19d7?source=rss-d9c8536f89ac------2</link>
            <guid isPermaLink="false">https://medium.com/p/b67d587a19d7</guid>
            <category><![CDATA[biology]]></category>
            <category><![CDATA[protein-engineering]]></category>
            <category><![CDATA[computational-biology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[biotechnology]]></category>
            <dc:creator><![CDATA[David de Meij]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 19:17:52 GMT</pubDate>
            <atom:updated>2025-11-02T19:17:52.904Z</atom:updated>
            <cc:license>http://creativecommons.org/licenses/by/4.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FhGWh3y8gAqTKJvi.jpg" /></figure><h3>From Evolution to Computation</h3><p>Proteins are the working parts of life. Every process in the body depends on proteins, whether repairing tissue, digesting food, or recognizing pathogens. For decades, biologists could only study what evolution had already produced, making small improvements through trial and error.</p><p>That changed when <a href="https://deepmind.google/science/alphafold/"><strong>AlphaFold2</strong></a> demonstrated that a protein’s 3D structure could be predicted directly from its amino acid sequence. It removed one of biology’s biggest bottlenecks. Now that we could predict how a protein folds, researchers began asking a new question: <em>what if we could design them from scratch?</em></p><p>Since then, models like <a href="https://github.com/RosettaCommons/RFdiffusion"><strong>RoseTTAFold Diffusion</strong></a> and <a href="https://generatebiomedicines.com/chroma"><strong>Chroma</strong></a> have expanded what’s possible. Instead of simply analyzing natural proteins, these systems can generate new ones with specific properties: enzymes that break down plastics, antibodies that bind tightly to disease targets, or entirely new molecular shapes that biology has never explored.</p><p>The shift is profound: biology is moving from being discovered to being engineered.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*wB0sKgbdCD8cqs5bJ37lrQ.png" /><figcaption>The AI-driven protein design loop: models generate candidate sequences, automated labs build and test them, and the results feed back to improve the next generation of designs.</figcaption></figure><h3>Two Philosophies of Design</h3><p>AI-based protein engineering has divided roughly into two camps.</p><ol><li><strong>Optimizing Nature.<br></strong> Companies like <a href="http://cradle.bio/"><strong>Cradle Bio</strong></a>, <a href="https://www.bighatbio.com/"><strong>BigHat Biosciences</strong></a>, and <a href="https://labgeniustx.com/"><strong>LabGenius</strong></a> focus on <strong>improving existing proteins</strong>. They use AI to predict which small mutations will enhance traits like binding strength, thermostability, or manufacturability. Their advantage lies in the data, which consists of millions of experimental measurements that feed back into their models. This approach builds on biology’s existing designs but makes the process faster and more targeted.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1MqAIQeD1BRaGsBE.jpg" /><figcaption>Illustration of Cradle’s ‘Optimizing Nature’ approach, where AI suggests small, data-driven mutations to improve existing proteins. Source: Cradle / TechCrunch</figcaption></figure><p><strong>2. Inventing Nature.<br></strong> Firms such as <a href="https://generatebiomedicines.com/"><strong>Generate:Biomedicines</strong></a>, <a href="https://www.absci.com/"><strong>Absci</strong></a>, <a href="https://www.profluent.bio/"><strong>Profluent</strong></a>, and <a href="https://www.xaira.com/"><strong>Xaira Therapeutics</strong></a> aim to create new proteins entirely (de novo). Their generative models start from a functional intent, such as “bind this receptor,” and then propose sequences that should fold into working structures. It’s a harder problem with higher failure risk, but the upside is enormous: completely new molecular tools unconstrained by evolution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*f52Zpk8rpcLA_cS_0jrfZQ.gif" /><figcaption>RFdiffusion generates a new protein designed to bind the insulin receptor, showing how generative models can create novel proteins from functional goals. Source: Ian C. Haydon / Baker Lab.</figcaption></figure><p>Both paths are valid. The first shortens the development cycle for industrial enzymes and drugs. The second expands the space of what biology can do.</p><p>Over time, these approaches will likely converge: in many cases design begins from reliable scaffolds, generative AI then explores novel sequence or structural space, and final candidates are refined through rapid experimental feedback.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dCeBLW82mjeeT_sx6kdFEg.png" /><figcaption><em>Overview of leading companies applying AI to protein engineering, grouped by whether they focus on inventing new proteins, optimizing existing ones, or combining both approaches.</em></figcaption></figure><h3>Platforms and Products</h3><p>The industry now divides into <strong>platform builders</strong> and <strong>product developers</strong>.</p><p>Platform builders such as <a href="http://cradle.bio/">Cradle</a>, <a href="https://www.helixon.com/">Helixon</a> and <a href="https://www.arzeda.com/">Arzeda</a> focus on developing general-purpose design systems. They train models on large experimental datasets and often license the resulting designs to partners. Their strength lies in robust data infrastructure and continuous learning loops.</p><p>Product developers such as <a href="https://generatebiomedicines.com/">Generate</a>, <a href="https://www.absci.com/">Absci</a>, and <a href="http://xaira.com/">Xaira</a> use these capabilities to build their own pipelines of therapeutics or materials. They control the full process: model training, lab validation, and clinical progression. This vertical integration allows them to capture more value if their candidates succeed but also exposes them to greater scientific and financial risk.</p><p>Most companies operate somewhere in between, providing design tools to partners while also advancing their own programs. It’s an ecosystem built around experimentation speed rather than a single dominant model.</p><h3>What Creates a Durable Advantage</h3><p>The most defensible companies share three things in common:</p><ol><li><strong>Closed feedback loops.<br></strong> Models that learn directly from in-house experimental results improve much faster than those trained only on public data. Integrating AI with automated labs, sometimes called “design-build-test-learn” cycles, creates a compounding advantage.</li><li><strong>Execution speed.<br></strong> The time between generating a design and validating it in the lab defines the pace of discovery. With automated synthesis and robotic assays, each step is logged and timed, turning experimentation into a measurable, optimizable process.</li><li><strong>Biological grounding.<br></strong> AI models can propose endless protein designs, but credibility comes from proving those designs work and are safe in real biology. Companies that can consistently demonstrate function and safety in cells or animals build proprietary datasets and know-how that others can’t easily replicate.</li></ol><p>Unlike in pure software, data and biology remain tightly coupled. Compute power helps, but the bottleneck is still generating reliable experimental results.</p><h3>Open-Source Foundations</h3><p>Open-source research has been essential to the rise of AI-driven protein design. It has allowed small academic groups and startups to experiment with powerful models that were once confined to major pharmaceutical companies. Unlike in software, however, most of these tools emerged <strong>after</strong> private breakthroughs rather than before them.</p><p>When <a href="https://github.com/google-deepmind/alphafold"><strong>AlphaFold2</strong></a> was released in 2021, it reshaped structural biology overnight. Though initially closed, its success inspired open reimplementations such as <a href="https://openfold.io/"><strong>OpenFold</strong></a> and <a href="https://www.ipd.uw.edu/2021/07/rosettafold-accurate-protein-structure-prediction-accessible-to-all/"><strong>RoseTTAFold</strong></a>, which soon matched much of its performance. Later, <a href="https://github.com/RosettaCommons/RFdiffusion"><strong>RFdiffusion</strong></a> and <a href="https://www.science.org/doi/10.1126/science.add2187"><strong>ProteinMPNN</strong></a> brought <em>de novo</em> design into the public domain, allowing researchers not just to predict structures but to create entirely new ones.</p><p>In parallel, non-ML open-source <strong>optimization and modeling</strong> frameworks — such as <a href="https://github.com/RosettaCommons/rosetta"><strong>Rosetta</strong></a>, <a href="https://rosettacommons.github.io/PyRosetta.notebooks"><strong>PyRosetta</strong></a>, and <a href="https://foldxsuite.crg.eu/"><strong>FoldX</strong></a> — provided the physical backbone for evaluating the stability, energetics, and docking behavior of designed proteins. More recent machine learning tools like <a href="https://github.com/gcorso/DiffDock"><strong>DiffDock</strong></a> and <a href="https://github.com/HannesStark/EquiBind"><strong>EquiBind</strong></a> extend this capability, predicting how AI-generated proteins might interact with ligands or other biomolecules. Together, these systems bridge data-driven design and molecular physics, ensuring that computationally imagined proteins remain chemically plausible.</p><p>A turning point came in 2024 with <a href="https://www.profluent.bio/modality/opencrispr"><strong>OpenCRISPR</strong></a>, developed by <a href="http://profluent.bio/">Profluent</a>. Using a closed generative model, the team designed a new CRISPR enzyme and released it freely to the public. The enzyme demonstrated around 95 percent fewer off-target edits than the standard SpCas9 and expanded access to genomic regions previously difficult to modify. By making the enzyme open-source, Profluent lowered barriers for academic labs often constrained by licensing costs or proprietary CRISPR systems.</p><p>Open and proprietary development now advance in parallel. Closed models focus on scale, integration, and proprietary data, while open projects emphasize transparency, reproducibility, and shared access. Together, they form a feedback loop that accelerates both discovery and democratization in molecular design.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JNuPi_g1-bSSdz8k-qA9pw.png" /><figcaption>Overview of major open-source tools underpinning AI-driven protein design.</figcaption></figure><h3>Where the Industry is Heading</h3><p>The near-term focus is on validation, which means demonstrating that AI-designed proteins can function as intended in clinical or industrial settings. By the late 2020s, it’s likely that the first therapeutics designed primarily by AI will complete human trials.</p><p>At the same time, industrial uses are scaling quietly in the background. For example, <a href="https://www.arzeda.com/post/arzeda-to-lead-major-nsf-funded-initiative-advancing-electricity-driven-cell-free-biomanufacturing-t"><strong>Arzeda</strong> designs enzymes for sustainable chemistry</a>. AI design enables companies to tailor proteins for stability, solubility, or specificity far faster than traditional screening.</p><p>Competition is intense, but the industry is gradually converging on hybrid pipelines:</p><ul><li><strong>Generative models</strong> to explore vast design space,</li><li><strong>Predictive models</strong> to evaluate candidates,</li><li><strong>Automated labs</strong> to test results and feed data back into training.</li></ul><p>This loop defines the new pace of biological innovation.</p><h3>The Broader Implication</h3><p>If protein design becomes reliable, it will change not just <strong>medicine</strong> but also <strong>manufacturing</strong>, <strong>food</strong>, and <strong>materials</strong>. Enzymes could replace chemical catalysts, new scaffolds could stabilize vaccines without refrigeration, and bio-based materials could outperform plastics.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b67d587a19d7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why the Future of Wind Energy is Airborne]]></title>
            <link>https://medium.com/@daviddemeij/why-the-future-of-wind-energy-is-airborne-f77d08abbeed?source=rss-d9c8536f89ac------2</link>
            <guid isPermaLink="false">https://medium.com/p/f77d08abbeed</guid>
            <category><![CDATA[renewable-energy]]></category>
            <category><![CDATA[offshore-development]]></category>
            <category><![CDATA[climate-change]]></category>
            <category><![CDATA[airborne-wind-energy]]></category>
            <category><![CDATA[wind-energy]]></category>
            <dc:creator><![CDATA[David de Meij]]></dc:creator>
            <pubDate>Wed, 05 Jan 2022 11:34:38 GMT</pubDate>
            <atom:updated>2022-08-06T09:54:54.902Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Airborne wind energy</strong> (AWE) is a form of wind energy where the object capturing the wind is lifted in the air by the wind itself tethered to the ground like a kite, instead of having blades attached to a large tower as with traditional wind turbines.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/787/0*KLFmxm1RG6YDUubB" /><figcaption>airborne wind energy system compared to a wind turbine (source: <a href="https://thekitepower.com/product/">Kitepower Falcon 100kW</a>).</figcaption></figure><h3><strong>Advantages of airborne wind energy</strong></h3><p>From a first principles standpoint airborne wind has a fundamental advantage: instead of having to build a heavy foundation and a large tower that can carry the load of the blades and turbine up in the air, all that is needed is a turbine (usually on the ground) and a winch with a tether that can handle the tension created by the kite.</p><p>This makes it possible to capture the wind at much <strong>greater heights</strong> where it is stronger and more reliable, while using much<strong> less material.</strong> At current scales up to 90% less material per MWh is used than traditional wind turbines [<a href="https://thekitepower.com/product/">Kitepower</a>], with larger scale systems this difference will only be more significant. There is a <a href="https://energyeducation.ca/encyclopedia/Wind_power">cubic relationship</a> between wind speed and the amount of energy that can be captured, meaning that if wind speed doubles the potential energy captured increases eight times. So capturing stronger winds at greater heights can drastically increase the amount of wind energy available.</p><h4>Improved logistics</h4><p>The smaller weight and smaller dimensions of airborne wind energy not only leads to lower material costs and a lower carbon footprint, it also <strong>reduces the logistics costs</strong> to deploy a system significantly. An airborne wind system can usually be stowed into one or more shipping containers and doesn’t have to be shipped using large specialized logistics and assembled onsite as for wind turbines. This makes airborne wind energy especially suitable for remote island locations, where shipping large wind turbines can be even more expensive.</p><h4>Reduced manufacturing &amp; operating costs</h4><p>Being smaller, it should also be cheaper and more achievable to<strong> mass manufacture</strong> these systems once commercialized on a larger scale. Furthermore, having the turbine on the ground instead of at a height <strong>reduces the maintenance cost</strong>, as you don’t need people climbing into a high tower which costs time and requires additional safety protocols.</p><h4>Wider range of wind speeds</h4><p>Another advantage is that an airborne system can operate in a<strong> wider range of wind speeds</strong>, increasing the amount of energy generated and also reducing the downtime making it more resemble baseload electricity. The kites can generally be safely landed and stowed away when a storm is predicted, making it less of a safety hazard.</p><h4>Deeper waters</h4><p>Airborne wind energy is also more interesting for <strong>deeper waters </strong>where current floating wind turbines are still too expensive at about <a href="https://www.energy.gov/sites/default/files/2021-08/Offshore%20Wind%20Market%20Report%202021%20Edition_Final.pdf">$160/Mwh</a> because they require massive floating foundations to balance the tower. Airborne wind only requires a relatively small floating platform anchored to the seabed.</p><h4>Noise and Visual pollution</h4><p>Finally, the <strong>noise </strong>and <strong>visual pollution</strong> of airborne wind energy is also expected to be reduced compared to traditional wind turbines. This hopefully makes local residents more positive towards airborne wind energy projects and limit the amount of opposition, making it potentially easier to get a permit. But this has to be seen in practice and also depends on the specifics of the system.</p><h3>So why don’t we see many airborne wind energy systems deployed already?</h3><p>Airborne wind energy is still a relatively young field of research, with its first patent issued in <a href="https://www.researchgate.net/publication/317327516_Comparing_Patent_and_Scientific_Literature_in_Airborne_Wind_Energy">1975</a> and the first scientific paper on tethered airfoils published in <a href="https://www.researchgate.net/publication/317327516_Comparing_Patent_and_Scientific_Literature_in_Airborne_Wind_Energy">1979</a>.</p><h4>Commercialization</h4><p>But with over 15 years of commercial development by various companies, some people may say that commercialization of airborne wind energy is long overdue. Luckily, in 2021 the industry seems to move towards <strong>commercializing the first systems</strong>. We have seen the first series production and delivery to a customer of an 80–200kW airborne wind energy system by <a href="https://skysails-group.com/blog/series-production-of-awes-starts/">Skysails</a> and Kitepower <a href="https://thekitepower.com/kitepower-in-aruba/">announced</a> the deployment of their 100kW system on Aruba this year as well. With more commercial and pilot projects from various companies expected in 2022 and beyond (i.e. <a href="https://www.kitemill.com/projects">Kitemill 5 20kW systems to be installed early 2022</a> with 12 EU-funded 100kW systems <a href="https://ec.europa.eu/clima/system/files/2021-12/policy_if_pf_2021_nawep_en.pdf">planned to be operational in 2024</a>, <a href="https://www.enerkite.de/news/332-neue-kooperation-mit-btu-cottbus-senftenberg-und-fraunhofer-pyco-zur-entwicklung-von-leichtbau-komponenten.html">Enerkite 100kW prototype</a> <a href="https://www.enerkite.de/news/332-neue-kooperation-mit-btu-cottbus-senftenberg-und-fraunhofer-pyco-zur-entwicklung-von-leichtbau-komponenten.html">expected to be completed in 2022</a>, <a href="https://www.ampyxpower.com/technology/demonstrator-ap3/">Ampyx Power 150kW AP-3</a> demonstrator <a href="https://www.ampyxpower.com/news/ampyx-power-2021-in-review/">expected to fly in 2022</a> or <a href="https://www.skypull.technology/solution">Skypull 100kW SP600 expected in 2024</a>).</p><h4>Airborne wind replaces atoms with bits</h4><p>One of the reasons that it has taken some time to get to this point, is that airborne wind energy systems are basically <strong>replacing atoms with bits</strong>. Instead of having a massive foundation and tower, we have a tethered kite or aircraft that has to be controlled through software. 10 years ago we might not have had the lightweight computer chips, sensors and the software needed to achieve this. Also, controlling a tethered aircraft or kite is a challenging technical problem with many potential points of failure.</p><h4>High material requirements</h4><p>Another reason for the prolonged development time is that airborne wind energy has <strong>high requirements for its materials</strong>. Both the tether and the kite has to be durable, lightweight and able to withstand high variable tensions for a long period of time.</p><h4>Automatically launching and landing</h4><p>Other challenges in airborne wind energy include <strong>automatically launching and landing</strong> the kite, which is a necessity to get to economic utility-scale power generation.</p><h4>More difficult to scale</h4><p>Airborne wind energy might also be <strong>more difficult to scale </strong>to a larger rated power. Since when you scale to a larger system all the material requirements also increase significantly, because the tether and kite need to withstand larger forces. As the kite gets larger and heavier it can be harder to maneuver and generates relatively less power (part of the lift force goes towards counteracting gravity).</p><h3>Different approaches</h3><p>There are currently many companies working on airborne wind energy with several different approaches.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*c-qpxA1iJpQeneEy" /><figcaption>Presently pursued concept (source: <a href="http://awesco.eu/awe-explained/#presently-pursued-concepts">AWESCO — AWE explained</a>).</figcaption></figure><h4>On ground versus on flying device electricity generation</h4><p>The first big difference in approaches is between:</p><ul><li>Generating electricity <strong>on the ground</strong> by using the lift power of the kite to unwind a winch on the ground connected to a generator.</li><li>Generating electricity<strong> in the air</strong> by using drag power to spin onboard rotors and using a conducting tether to send the electricity down.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Tnlrk9zrvVgIglFn" /><figcaption>Lift power versus drag power (source: <a href="http://awesco.eu/awe-explained/#historical-perspective">AWESCO — AWE explained</a>).</figcaption></figure><h4>On-flying-device electricity generation</h4><p>The main advantage of generating the electricity <strong>on the flying device</strong> is that you have continuous power generation in contrast to a pumping cycle operation which uses some of the generated energy to reel the kite back in. The main disadvantages are that having the generators in the air increases the weight of the flying device and also introduces additional complexity.</p><p><a href="https://x.company/projects/makani/">Makani</a> (acquired by Google X) put a lot of time and resources into this approach but was discontinued after 13 years of development because “<a href="https://techcrunch.com/2020/02/18/alphabet-takes-the-wind-out-of-its-makani-energy-kites/">the road to commercialization is longer and riskier than hoped</a>”. Alphabet did publish a <a href="https://storage.googleapis.com/x-prod.appspot.com/files/Makani%20Non-Assertion%20Pledge.pdf">non assertion pledge</a> for the over 400+ patents and applications of Makani and carefully documented what had been developed and what can be improved, also providing all their software with an <a href="https://github.com/google/makani">open source license</a>. Makani also made an interesting documentary <a href="https://www.youtube.com/watch?v=qd_hEja6bzE&amp;feature=youtu.be">Pulling Power from the Sky: The Story of Makani</a>. All of this is a wealth of information for the start-ups such as <a href="https://www.kitekraft.de/#Kite">Kitekraft</a> and <a href="https://windlift.com/">Windlift</a> that are working on a similar approach.</p><h4>On-ground electricity generation</h4><p>The area where currently most commercial development seems to happen, is airborne wind energy systems with a <strong>fixed ground station </strong>with the kites <strong>flying crosswind</strong>, usually in patterns of 8 using a “pumping cycle”. This means first unwinding a winch to generate electricity, after which the kite is reeled back in (using a small fraction of the generated electricity).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fj1pomHgHKiFkwi1vhHQWg.gif" /><figcaption>Insightful animation illustrating the “pumping cycle” flight path (source: <a href="https://thekitepower.com/product/">Kitepower Falcon 100kW system</a>).</figcaption></figure><h4>Kite turbines</h4><p>There are also companies doing ground-based electricity generation by using rotating kites or blades that power a rotor on the ground. For example, <a href="https://windswept-and-interesting.co.uk/">Windswept-and-interesting</a> (new site: <a href="http://windswept.energy/">windswept.energy</a>) is developing “<strong>Kite Turbines</strong>”, a concept best explained in this recent <a href="https://www.bbc.com/news/av/uk-scotland-59190335">item by the BBC</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/0*dBnX6hZWLCPA9Gvx.png" /><figcaption><a href="https://windswept-and-interesting.co.uk/">Windswept</a> kite turbines involve multiple spinning kites to drive a generator</figcaption></figure><p>This concept is currently still in an early stage, being tested on a small scale with not many companies or academics pursuing this option currently.</p><p>Windswept <a href="https://github.com/rodread/DaisyKiteTurbineControl">open-sourced</a> their idea for &lt;1 kW kite turbines. This might be useful for <strong>remote villages </strong>in the<strong> developing world</strong>. As it is a relatively<strong> </strong>simple system that can be built with cheap materials.</p><h4>Rigid wing versus soft fabric</h4><p>Among the companies working on airborne wind energy, one major difference is whether to use <strong>rigid wing </strong>aircraft, <strong>soft fabric </strong>kites or something in between. Advantages of a rigid wing may be better aerodynamics (which means more energy generated for the same size kite) and better durability. Advantages of a soft fabric kite are probably cost of material and manufacturing, safety (as a soft fabric kite crashing will most likely cause less damage) and easier to ship (as a soft fabric kite is easier to stow away).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t7jkzrzRnVa-xjkCugma7g.png" /><figcaption>From left to right: <a href="https://skysails-power.com/onshore-units/">Skysails</a> soft fabric kite, <a href="https://www.enerkite.de/technologie.html">Enerkite</a> hybrid lightweight wing with rigid frame, <a href="https://www.ampyxpower.com/future/products-and-markets/">Ampyx Power</a> rigid body AP4 aircraft.</figcaption></figure><h4>Launching and landing methods</h4><p>Then there is also a difference in how different companies <strong>launch </strong>and<strong> land</strong> their kites. For example, Ampyx Power is planning to build a special launching and landing pad using a <a href="https://www.ampyxpower.com/news/ampyx-power-and-nlr-join-forces-to-develop-alternative-wind-energy-solutions/">catapult system</a>, Skysails uses a <a href="https://skysails-power.com/how-power-kites-work/">launch and landing mast</a>, Enerkite uses a <a href="https://www.enerkite.de/technologie.html">rotating telescopic mast</a> to launch the kite, Skypull uses a <a href="https://www.skypull.technology/tech">“box-wing” drone</a> for vertical take-off and landing and other companies such as <a href="https://thekitepower.com/product/#faq">Kitepower</a> currently rely on training people how to launch the kite and use a special flight path to keep the kite in the air in low wind conditions to avoid landing and relaunching.</p><h3>The future of airborne wind energy</h3><p>Unlike with traditional wind turbines where there is not much variation between companies, there are currently <strong>many different approaches</strong> in airborne wind energy. Airborne wind energy systems with a <strong>soft body kite</strong> and <strong>on-ground energy generation</strong> are the first to commercialize. But for rigid wing systems, there seems to be a more defined roadmap for scaling to larger systems (at least from publicly available information). For example the 2MW+ <a href="https://www.ampyxpower.com/future/products-and-markets/">Ampyx Power AP4</a>, the 500kW <a href="https://enerkite.de/en/#produkte">Enerkite EK1M</a>, the 1MW <a href="https://www.skypull.technology/solution">Skypull SP1700</a> or the 3MW <a href="https://www.kitemill.com/projects">Kitemill KM5</a>.</p><p>As more airborne wind energy systems get commercialized, we will find out which approaches are the <strong>most economical </strong>and <strong>scalable</strong> and I expect the industry to converge on one approach at least for utility-scale electricity generation (there may exist different approaches for specific niches). The companies that are too slow to commercialize their technology or are focussed on the wrong approach without adapting quickly, might not survive the end of this decade as the airborne wind energy industry matures.</p><p>Once airborne wind energy systems have proven themselves on land, they will most likely quickly move to offshore applications. Especially to <strong>deeper waters</strong> where current wind turbines are challenging and expensive and to <strong>decommissioned offshore wind farms</strong>, where the existing infrastructure can readily be used.</p><p>Once the reliability and safety of airborne wind energy has been proven, we might also see it <strong>on top of apartment buildings</strong> and <strong>skyscrapers</strong> in dense cities as an additional locally produced power source that can work in conjunction with solar energy.</p><h3>Conclusion</h3><p>The possibility of cheap renewable electricity that airborne wind energy can provide is very promising for its potential for accelerating the transition to renewable energy and for its potential of <strong>alleviating energy poverty </strong>for a large part of the world that currently does not benefit from the luxury of affordable electricity. It is going to be very exciting to see how the airborne wind energy industry evolves over the coming years.</p><p>If you have any remarks or find any inaccuracies in this article, please let me know at <a href="/service/mailto:daviddemeij@gmail.com">daviddemeij@gmail.com</a>.</p><h3><strong>Additional resources</strong></h3><ul><li><a href="http://awesco.eu/awe-explained/">AWESCO — AWE Explained</a> a Great technical explanation of airborne wind and different approaches.</li><li><a href="https://airbornewindeurope.org/about-airborne-wind-energy/">Airborne Wind Europe</a> is the association of the European airborne wind energy industry.</li><li><a href="https://thekitepower.com/product/">Kitepower — Product</a> great explanation of their Falcon 100kW system.</li><li><a href="https://skysails-power.com/how-power-kites-work/">Skysails — How it works</a> visual overview of the components of their 80–200kW system.</li><li><a href="https://www.skypull.technology/tech">Skypull — Technology</a> innovative “box-wing” drone concept.</li><li><a href="http://windswept.energy/">Windswept-and-Interesting</a> lightweight “kite turbine” concept.</li><li><a href="https://www.ampyxpower.com/future/products-and-markets/">Ampyx Power — Products and markets</a> roadmap to a 5MW system.</li><li><a href="https://www.kitemill.com/projects">Kitemill — Projects</a> roadmap to a 3MW system.</li><li><a href="https://enerkite.de/en/#produkte">Enerkite — Product</a> roadmap to a 500kW system.</li><li><a href="https://www.skypull.technology/solution">Skypull — Solution</a> roadmap to a 1MW system</li><li><a href="https://medium.com/kitekraft">Kitekraft — Blog</a> insightful articles on airborne wind energy from the founder of Kitekraft.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f77d08abbeed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How nbdev helps us structure our data science workflow in Jupyter Notebooks]]></title>
            <link>https://medium.com/overstoryai/how-nbdev-helps-us-structure-our-data-science-workflow-in-jupyter-notebooks-9cf6081b051f?source=rss-d9c8536f89ac------2</link>
            <guid isPermaLink="false">https://medium.com/p/9cf6081b051f</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[jupyter-notebook]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[David de Meij]]></dc:creator>
            <pubDate>Wed, 15 Jan 2020 14:44:25 GMT</pubDate>
            <atom:updated>2020-11-10T11:34:56.882Z</atom:updated>
            <content:encoded><![CDATA[<p>Jupyter notebooks have been rightly praised for making it very easy and intuitive to experiment with code, visualize results and describe your process in nicely formatted markdown cells. In our work as data scientists and Machine Learning Engineers at <a href="https://overstory.ai/">overstory.ai</a> (formerly: <a href="https://www.20tree.ai/">20tree.ai</a>), we are using notebooks all the time and they are a great tool in our toolbox. However, there are some downsides that we experience using notebooks in our workflow. For one thing, working in notebooks often results in unstructured, badly documented and untested code that takes a lot of time to transport into a proper codebase. Also, version control in Jupyter Notebooks is a disaster and involves large changes in each commit that are hard to traceback.</p><p><a href="https://nbdev.fast.ai/">Nbdev</a> solves many of these issues by making it very easy to transform Jupyter notebooks into proper python libraries and documentation. It incentives us to write clear code, use proper Git version control and document and test our codebase continuously. At the same time preserving the benefits of having interactive Jupyter notebooks in which it is easy to experiment. Using nbdev has improved our workflow significantly and perhaps it can do the same for you!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*TJ-lMmRoMOPU70ZLj0fbyw.gif" /></figure><h3>What is nbdev</h3><p><a href="https://nbdev.fast.ai/">nbdev</a> is a library created by <a href="https://fast.ai/">fast.ai</a> that forms the missing link between the data exploration of Jupyter notebooks and programming an actual codebase that produces high-quality software. In short, nbdev provides a framework for:</p><ul><li>Exporting selected notebook cells into a Python library</li><li>Automatically generating documentation based on function signatures and notebook cells</li><li>Running notebook cells as unit tests.</li></ul><p>Of course, it has many more features that you can find at <a href="http://nbdev.fast.ai/">nbdev.fast.ai</a> or by reading the <a href="https://www.fast.ai/2019/12/02/nbdev/">nbdev launch post</a>, but these are the ones we make the most use of. In addition, nbdev provides utilities for stripping notebook metadata and handling merge conflicts, both greatly improving development cycles (and general quality of life :) ) when checking notebooks into a git repository.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/229/1*vqiOC9nBA7MNGGu9WPpyjg.png" /><figcaption>Every notebook that has a <strong>#default_exp </strong>tag at the start, gets exported to a Python module. Automatically using best practices such as defining which functions get imported when you do <strong>import *</strong> by setting<strong> __all__ = [‘function1’, ‘functon2’, …]</strong>.</figcaption></figure><p>The magic of nbdev is that it doesn’t actually change programming that much; you add a # export or # hide tag to your notebook cells once in a while, and you run nbdev_build_lib and nbdev_build_docs when you finish up your code. That’s it! There’s nothing new to learn, nothing to unlearn. It’s just notebooks. More than anything, nbdev encourages us to write <strong>cleaner</strong> notebooks with a <strong>clear separation</strong> between <strong>code</strong> (the cells that get exported) and <strong>experimentation, visualization and testing</strong> (the cells that don’t get exported). As a result, our notebooks are more readable, and easier to share among the whole team. As a wise (wo)man once said, the real nbdev was inside of us all along.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cyavwHSMnX8IVp6O.png" /><figcaption>index.ipynb with the generated documentation from this notebook next to it.</figcaption></figure><h3>Why do we use it?</h3><p>As data scientists, much of our work involves — as you might expect — data. This means loading data, transforming data, combining data, and at some point actually using that data. Especially in the transforming and combining stages, it’s critical to ensure that no mistakes slip in. If you are trying to train a neural network for semantic segmentation but your segmentation map is shifted by a few pixels, your data is essentially invalid. Worse yet, valuable time is often lost trying to get some code or machine learning models to work, while minor typos have sneaked in such small, stupid, but bothersome bugs. Because notebooks are run iteratively through cells it’s almost like you’re debugging <em>while</em> coding. Errors are much quicker caught this way. Here at <a href="https://20tree.ai/">20tree.ai</a>, we mostly work with georeferenced (satellite) data. When different parts of your data are projected in different <a href="https://en.wikipedia.org/wiki/Spatial_reference_system">coordinate reference systems</a>, it (again) becomes very easy for mistakes to slip in.</p><p>When developing new code, a pretty standard pattern for us consist of the following:</p><ol><li>Small functions are written in a Jupyter notebook. The notebook is used to visually inspect the output and to informally test that the code behaves as expected;</li><li>The functions get copy-pasted into a proper codebase;</li><li>The original notebooks are scattered to the wind;</li><li>Code gets changed over time, maybe a mistake slips in. When asking for more details on a bit of code, someone points to some file called Untitled_v3_better_labels2.ipynb with the comment “it’s probably very outdated though”.</li></ol><p>With nbdev, we can nip this whole sequence in the bud. You write your code in a Jupyter notebook and that’s it. You’re done, because the notebook <em>is</em> the proper codebase! The main code cells are exported to the library and the output of some cells forms the visual explanation as well as unit tests — that you’d have to make generally separately otherwise — that are automatically run when pushing the notebook to Github. So while iteratively coding documentation and testing are (almost) entirely free. For example, after implementing data augmentation, you’re surely going to visualize the outputs to ensure that the data looks as you would expect and, just as importantly, that the labels are similarly transformed. With nbdev, this visualization is simply in your codebase, right below the function definition:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/944/0*Z44YB8wGXZPlyS1-" /></figure><p>These tests can also very easily be used in your Github continuous integration pipeline, making it very easy to do proper checks before merging some new code into your existing codebase.</p><p>The fact that your entire codebase is living in notebooks also means that when something is not working as expected, it is very easy and intuitive to debug. You can quickly change something and see how it affects the output.</p><h3>Best practices we found</h3><p>In the past month that we have been using nbdev in our workflow, we have developed some best practices.</p><ul><li>Use test flags such as server, cuda or slow that can be set in your settings.ini by setting tst_flags = server|cuda|slow. Cells that are given this flag will not be run during testing unless specifically requested. This can be useful if, for example, you are running a function using a large dataset that is only available on the server.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/667/1*gPRon4var07QdOVAAX5amw.png" /><figcaption>Add a <strong>#server</strong> the start of the cell to avoid it being run during testing.</figcaption></figure><ul><li>Be extra careful to make clean commits as changes in Jupyter Notebook are not very readable on Github. The nbdev command-line functions nbdev_fix_merge (which splits merge conflicts) and nbdev_clean_nbs (which removes metadata from your notebooks) are very useful for this. Also, we use git add -p to select specific changes in a notebook to be added to the commit. If you are using Jupyter Lab you can also install the <a href="https://github.com/jupyterlab/jupyterlab-git">Git extension</a> to view changes.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/631/1*WK-ii-42G49MvtljL04Kwg.png" /><figcaption>You can also use the Jupyter Lab <a href="https://github.com/jupyterlab/jupyterlab-git">Git extension</a> to directly view changes in your Notebooks or you can use the Github <a href="https://www.reviewnb.com/">ReviewNB</a> app to view all the changes for a particular PR (paid for private repositories).</figcaption></figure><ul><li>Always add some sanity checks after defining a function, since it is automatically added as a test this makes your code much more reliable, it also shows the user the expected output in the documentation. If you don’t want a test to appear in your documentation you can simply add #hide at the beginning of the cell (the test will still be run). Use nbdev_test_nbs to test all the Notebooks locally.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*-IXAQggwpVivLw5Be3XloA.gif" /><figcaption>You can use <strong>test_eq()</strong> to quickly run a test while at the same time showing the expected output in the docs.</figcaption></figure><ul><li>For defining command-line functions, we use <a href="https://github.com/fastai/fastscript">fastscript</a>, another great library by <a href="https://fast.ai/">fast.ai</a>, which makes it possible to define a command-line function that can also be run as a regular function, making it very convenient to test in the notebook as well. You can add your command line functions by adding them in settings.ini (separated by spaces).</li></ul><pre>console_scripts = math_add=math.core:add_floats</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/791/1*d6gixnNf7cdk8Ok0nUIC4Q.png" /></figure><ul><li>Many of the awesome widgets that make Jupyter Notebooks so powerful can be displayed in the documentation, as long as it can be rendered as an HTML output (which is possible for most widgets). For example, we can use folium to render a map with a GeoJSON layer in our documentation.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tENhPC5tng93irdsHww3OQ.gif" /><figcaption>We can even include widgets in our nbdev documentation, as long as they give an HTML output.</figcaption></figure><h3>Authenticated nbdev docs</h3><p>The only real problem we encountered with using nbdev for generating our documentation is that some of our documentation contains sensitive customer data and other information that we want to keep secure, which means that we don’t want to host it on a public Github pages website.</p><p>To solve this issue we use <a href="https://github.com/benbalter/jekyll-auth">https://github.com/benbalter/jekyll-auth</a> which makes it very easy to host a Jekyll website on Heroku using Github authentication. The beauty of this solution is that you can immediately give your entire Github organization or Github team access to the documentation. For the full details on implementing this, check out our other blog post on <a href="https://medium.com/20tree-ai/auto-generated-authenticated-python-docs-using-nbdev-and-heroku-49c7f29ddd72">Auto-Generated Authenticated Python Docs using nbdev and Heroku</a>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/646/1*vkh09AvFxiG2uwreUgzr9Q.png" /><figcaption>Using Github OAuth for granting access to our internal documentation.</figcaption></figure><h3>Conclusion</h3><p>Even though we only just started using nbdev, it has already proven to be extremely useful to us. The idea of building libraries entirely from within Notebooks is truly paradigm-shifting and makes us not only more effective both individually and as a team, but also allows us to be much more inter-disciplinary, through feature-rich, visual, and well-documented code.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9cf6081b051f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/overstoryai/how-nbdev-helps-us-structure-our-data-science-workflow-in-jupyter-notebooks-9cf6081b051f">How nbdev helps us structure our data science workflow in Jupyter Notebooks</a> was originally published in <a href="https://medium.com/overstoryai">Overstory</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Auto-Generated Authenticated Python Docs using nbdev and Heroku]]></title>
            <link>https://medium.com/overstoryai/auto-generated-authenticated-python-docs-using-nbdev-and-heroku-49c7f29ddd72?source=rss-d9c8536f89ac------2</link>
            <guid isPermaLink="false">https://medium.com/p/49c7f29ddd72</guid>
            <category><![CDATA[fastai]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[jupyter-notebook]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[David de Meij]]></dc:creator>
            <pubDate>Wed, 15 Jan 2020 14:42:51 GMT</pubDate>
            <atom:updated>2020-01-15T14:42:51.292Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://nbdev.fast.ai/">Nbdev</a> is a python library developed by <a href="https://fast.ai/">fast.ai</a> that makes it very convenient to create libraries and documentation from <a href="https://jupyter.org/">Jupyter Notebooks</a>. To learn more about nbdev and it’s applications check out the documentation at <a href="https://nbdev.fast.ai/">nbdev.fast.ai</a> (in a brilliant recursive manner generated with nbdev itself) or read our <a href="https://medium.com/20tree-ai/how-nbdev-helps-us-structure-our-data-science-workflow-in-jupyter-notebooks-9cf6081b051f">other blog post</a> on how and why we have been using nbdev in our team, how it has improved our workflow and what the best practices are that we have found.</p><p>One of the problems we encountered with using nbdev for generating our documentation is that some of our documentation contains sensitive customer data and other information that we want to keep secure, which means that we don’t want to host it on a public Github pages website.</p><p>To solve this issue we use <a href="https://github.com/benbalter/jekyll-auth">jekyll-auth</a> which makes it very easy to host a Jekyll website on Heroku using Github authentication. The beauty of this solution is that you can immediately give your entire Github organization or Github team access to the documentation.</p><p>Making your nbdev documentation an authenticated Heroku app is done very quickly following a few simple steps.</p><h4><strong>Step 1. Create an Heroku App</strong></h4><p>Install Heroku toolbelt from <a href="https://devcenter.heroku.com/articles/heroku-cli">https://devcenter.heroku.com/articles/heroku-cli</a> and run:</p><pre>heroku create appname</pre><p>Copy the generated URL (i.e. <a href="https://myapp.herokuapp.com/).">https://appname.herokuapp.com/).</a></p><figure><img alt="note that I had to use mycoolapp20 since mycoolapp was already taken on Heroku." src="https://cdn-images-1.medium.com/max/690/1*PnNZO9RVWPL0h66hnnrVbg.png" /><figcaption>Note that I used <strong><em>mycoolapp20</em></strong> because <strong>mycoolapp</strong> was already taken on Heroku.</figcaption></figure><h4><strong>Step 2. Create an OAuth App in Github</strong></h4><p>Go to <a href="https://github.com/organizations/[org_name]/settings/applications/">https://github.com/organizations/[org_name]/settings/applications/</a> and create a new OAuth App and set the homepage URL to <a href="https://appname.herokuapp.com/">https://appname.herokuapp.com/</a> and the callback URL to <a href="https://appname.herokuapp.com/auth/github/callback">https://appname.herokuapp.com/auth/github/callback</a> using the same appname as earlier (<strong>important</strong>). Keep the tab with the generated CLIENT_ID and CLIENT_SECRET open for later use.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/522/1*SQL6Dr4W5lV49VGs5lOUTA.png" /></figure><h4>Step 3. Upload to Heroku</h4><p>For some reason, when hosting the nbdev docs on Heroku the URLs will only work when there is no host or base URL set. So, before building your docs make sure to set <em>doc_host</em> and<em> doc_base </em>to<em> None</em> in your app’s nbdev settings.inias follows:</p><pre>doc_host = <br>doc_base = </pre><p>Then to build the nbdev docs run:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/55820d9287669645b214e768b98d08a3/href">https://medium.com/media/55820d9287669645b214e768b98d08a3/href</a></iframe><p>Then install jekyll-auth:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/1dec450e8f862cd374dc9058123efba7/href">https://medium.com/media/1dec450e8f862cd374dc9058123efba7/href</a></iframe><p>Now set the CLIENT_ID<em> </em>and CLIENT_SECRET<em> </em>from your Github OAuth application in Heroku. Also adding either you<em>r </em>GITHUB_ORG_NAMEas written in your Github Organisation URL (i.e. 20treeAI in our case <a href="https://github.com/organizations/20treeAI/s">https://github.com/organizations/20treeAI/</a>) or your GITHUB_TEAM_ID (to find you Team ID check the instructions at the bottom of the Jekyll-Auth <a href="https://github.com/benbalter/jekyll-auth/blob/master/docs/getting-started.md">Getting Started</a> doc).</p><pre>heroku config:set GITHUB_CLIENT_ID=[CLIENT_ID] GITHUB_CLIENT_SECRET=[CLIENT_SECRET] GITHUB_ORG_NAME=[ORG_NAME] # or GITHUB_TEAM_ID=[TEAM_ID]</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tjv0XlD6tmWUXUDRYJtzDQ.png" /></figure><p>Then install and push to Heroku using:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/20433560ea6fb39acee61344b2efad29/href">https://medium.com/media/20433560ea6fb39acee61344b2efad29/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lqABxVnA4JuElmoc3OA3yQ.png" /></figure><p>You should now be able to visit the website typing heroku open in the terminal.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/646/1*vkh09AvFxiG2uwreUgzr9Q.png" /><figcaption>The first time it asks you to allow the app to verify your Github organization or team.</figcaption></figure><p>After allowing access you should be able to see your documentation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yRl6K46b1C3db4rkqTEgzA.png" /><figcaption>index.ipynb with the generated documentation from this notebook next to it.</figcaption></figure><h4>Step 4. Updating the Docs</h4><p>Then to update the docs on Heroku run the following code from the <em>app</em> directory (not from docs):</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/53c559226d8d802afc7719c28637297e/href">https://medium.com/media/53c559226d8d802afc7719c28637297e/href</a></iframe><p>If for some reason you deleted your Heroku remote, you can run the following code before pushing to Heroku.</p><pre>heroku git:remote -a appname</pre><p>Or if you have a merge conflict for example because of unrelated git histories (this can happen if you squash commits locally but not on Heroku) or because the local branch is somehow behind, you can reset the repo before pushing to Heroku.</p><pre>heroku repo:reset</pre><p>This is how we have been using <a href="https://github.com/benbalter/jekyll-auth">jekyll-auth</a> to make our nbdev documentation easily accessible to our team members. Hopefully, this will be useful to other teams around the world.</p><h4><strong>Other Resources</strong></h4><ul><li><a href="https://nbdev.fast.ai/">nbdev.fast.ai</a></li><li><a href="https://github.com/benbalter/jekyll-auth">Jekyll-Auth</a> <a href="https://github.com/benbalter/jekyll-auth/blob/master/docs/getting-started.md">Getting Started</a></li><li><a href="https://medium.com/20tree-ai/how-nbdev-helps-us-structure-our-data-science-workflow-in-jupyter-notebooks-9cf6081b051f">How nbdev helps us structure our data science workflow in Jupyter Notebooks</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=49c7f29ddd72" width="1" height="1" alt=""><hr><p><a href="https://medium.com/overstoryai/auto-generated-authenticated-python-docs-using-nbdev-and-heroku-49c7f29ddd72">Auto-Generated Authenticated Python Docs using nbdev and Heroku</a> was originally published in <a href="https://medium.com/overstoryai">Overstory</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>