<?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[CodeElevation - Medium]]></title>
        <description><![CDATA[Think deeper. Code smarter. Build better. - Medium]]></description>
        <link>https://medium.com/codeelevation?source=rss----1c377a233996---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>CodeElevation - Medium</title>
            <link>https://medium.com/codeelevation?source=rss----1c377a233996---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 02:45:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/codeelevation" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com" rel="hub"/>
        <item>
            <title><![CDATA[My Model Was Accurate, but My Machine Learning System Was Still Wrong]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/my-model-was-accurate-but-my-machine-learning-system-was-still-wrong-bbeb9fb7a12c?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/1672/1*PzVe3CKYHplAJ-sQZMv76A.png" width="1672"></a></p><p class="medium-feed-snippet">I spent too much time optimizing model metrics before realizing that accuracy was only measuring one small part of whether the system&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/my-model-was-accurate-but-my-machine-learning-system-was-still-wrong-bbeb9fb7a12c?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/my-model-was-accurate-but-my-machine-learning-system-was-still-wrong-bbeb9fb7a12c?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/bbeb9fb7a12c</guid>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Maximilian Oliver]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 05:28:36 GMT</pubDate>
            <atom:updated>2026-11-18T05:28:35.500Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Protecting a System That Can’t Say No]]></title>
            <link>https://medium.com/codeelevation/protecting-a-system-that-cant-say-no-d14c34f036d2?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/d14c34f036d2</guid>
            <category><![CDATA[legacy-systems]]></category>
            <category><![CDATA[rate-limiting]]></category>
            <dc:creator><![CDATA[Adwait Mathkari]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 05:21:00 GMT</pubDate>
            <atom:updated>2026-11-18T05:20:59.581Z</atom:updated>
            <content:encoded><![CDATA[<h3>Rate limiting a legacy ERP from the consumer side — because it cannot shed load on its own.</h3><p>Most of the interesting engineering in enterprise software happens at the seams — where something modern talks to something that has been running the business for twenty years.</p><p>In our case the modern thing is a Node.js GraphQL backend serving a B2B customer portal. The old thing is an ERP system, and every meaningful action a customer takes eventually becomes an HTTP call to it.</p><p>This article has two halves. The first is the business problem and the rules we settled on. The second is the engineering problem sitting underneath them, which turned out to be far more interesting than counting requests.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ptn9cXBQpEBlLMDJv06ClA.png" /></figure><h3>Part one: the business problem</h3><p>The ERP is a legacy system, and it is not elastic. Capacity is expensive and is whatever was provisioned years ago; it does not grow under pressure. It has a finite pool of work processes, and when that pool is exhausted it does not gracefully shed load. It slows down. Then it slows down more. Then requests start timing out.</p><p>Here’s the vicious part. When a user sees a spinner that never resolves, they hit refresh. Now the same query is in flight twice — and the person most likely to refresh again is the one already waiting. The system spirals into a state it cannot recover from.</p><p>The ERP isn’t ours alone, either. Other applications depend on it. If our portal saturates the shared backend, we don’t just break our own product. We take down someone else’s.</p><p>So the problem statement was never “make the portal faster.”</p><blockquote><em>Our application must never be the reason the ERP falls over. If we cannot serve a request safely, we must refuse it quickly, cheaply, and honestly.</em></blockquote><p>Refusing work <strong>is</strong> the feature. A fast, clear 429 is better for everyone than a request that sits in a queue for 55 seconds, times out, and gets retried.</p><h3>Choosing what to measure</h3><p>We had three candidate units on the table, and the choice between them shaped everything downstream.</p><p><strong>Arrival rate — requests per second, or per minute.</strong> The default. Well understood, supported by every library, trivially explainable to anyone who asks. Genuinely a reasonable marker, and for a backend with predictable service times it is the right answer. Window size barely changes the picture either: a minute is coarse enough that an entire burst can arrive, saturate the ERP and time out inside one window; a second is finer, but it measures the same thing.</p><p><strong>Concurrent in-flight requests.</strong> The ERP serves one request with one work process, held for the whole duration of the call. So a count of our in-flight requests is a direct count of the threads we are occupying — the same finite pool everything else is competing for. Unconventional as a limiter, no off-the-shelf implementation, harder to explain.</p><p><strong>Observed backend health.</strong> Not a measurement of our traffic at all, but of how the ERP is responding to it — P90 latency, and the share of recent calls that timed out.</p><p>We discarded the first and used the other two together.</p><p>An arrival-rate cap controls how fast work <em>enters</em>. It says nothing about how much work is <em>resident</em> — and the two only correspond when service time is stable. Ours isn’t. Service time is the first thing that moves when the ERP degrades, and it moves by an order of magnitude. A cap of 5 requests per second means 10 concurrent requests on a healthy day and 250 on a bad one. The limit is loosest exactly when you need it tightest.</p><p>Concurrency, by contrast, is the thing the ERP actually experiences. Health is the thing it reports back. One measures what we are asking for; the other measures whether it can cope.</p><p>Two units, two families of rules.</p><h3>Family one: concurrency ceilings</h3><p>Counters of what is happening <em>right now</em>, at five scopes, each with its own threshold.</p><p><strong>Global</strong> — total requests in flight to the ERP.</p><p><strong>Per endpoint</strong> — how many of the known-expensive calls are in flight.</p><p><strong>Per large customer</strong> — is one high-volume account monopolising an expensive endpoint?</p><p><strong>Per customer</strong> — is one account monopolising the backend generally?</p><p><strong>Per user</strong> — is one person hammering F5?</p><p>Cross a threshold and the request is refused immediately with a 429.</p><h3>Family two: health signals</h3><p>Ceilings protect against volume. But the ERP can be unhealthy while our traffic is perfectly normal — a batch job, a bad query plan somewhere else, an infrastructure problem.</p><p>So we also watch the <em>behaviour</em> of recent calls over a rolling five-minute window, and collapse it into a single isSlow verdict.</p><p><strong>Timeout ratio</strong> — if 30% or more of recent calls timed out, the ERP is considered slow.</p><p><strong>P90 latency</strong> — if the 90th-percentile duration crosses roughly 50 seconds, same verdict.</p><p>That difference in what the two families measure reaches all the way out to the API contract. A ceiling rejection means <em>“you, specifically, are asking for too much right now.”</em> An isSlow rejection means <em>&quot;the backend itself is degraded.&quot;</em> The frontend shows a different message for each, because the user needs to do something different: try again shortly, versus the system is having a bad day.</p><h3>One chokepoint, not many</h3><p><strong>There is exactly one place where all of this is enforced.</strong></p><p>Not in resolvers. Not sprinkled through datasources. Every call to the ERP funnels through a single adapter method — and the guard lives there.</p><p>If someone adds a new query next year and forgets rate limiting exists, they are still protected. Enforcement you have to remember to apply isn’t enforcement. It’s a convention.</p><h3>Part two: the engineering problem</h3><p>Everything above is a specification. The interesting part is what it takes to enforce it correctly when the application doing the enforcing is horizontally scaled and asynchronous.</p><h3>The counter is shared state</h3><p>Anyone who has written a rate limiter has met this problem on day one: the count has to live somewhere every request can see, and the check against it has to be atomic. It isn’t a twist at the end of the story. It’s the first thing you deal with.</p><p>Still, the Node.js version of it is sharper than most, and worth stating precisely.</p><pre>// This looks correct. It is not.<br>const active = await store.get(&#39;activeRequests&#39;);<br>if (active &gt;= LIMIT) {<br>  throw new TooManyRequests();<br>}<br>await store.set(&#39;activeRequests&#39;, active + 1);</pre><p>Read, compare, write — three steps with an await between them. In Node an await is a yield, not a pause.</p><p>When a burst of requests lands in the same tick, each one issues its read and hands control straight back to the event loop. Every read in the burst goes out before any response comes back. So they all read the same value — zero — all conclude they are under the limit, and all proceed. The limit is never applied.</p><p>This is an ordinary read-modify-write race. What makes it worse here is the size of the window: the gap between read and write is a full network round trip, and the event loop will cheerfully admit every other pending request into that gap.</p><p>It also isn’t rare or load-dependent. It is guaranteed under precisely the burst conditions the limiter exists for — a dashboard mounting and firing eight queries at once, across a few hundred users, at 9am.</p><p>So the check and the increment have to be one indivisible step, and every request has to pass through a single point that serialises them. The counter climbs 1, 2, 3 … 300, and each request sees the value <em>including its own increment</em> before deciding whether it&#39;s allowed.</p><p>That leaves one question: where does that single point live?</p><h3>Redis: the single thread we already had</h3><p>The obvious candidates don’t survive contact with the deployment.</p><p>You can’t keep the counter in process memory. There are multiple Node instances behind a load balancer, and a per-instance counter protects nothing — five instances with a limit of 300 is a limit of 1,500.</p><p>You can’t put a real message queue in front of it without turning a synchronous GraphQL request into an async workflow — which changes the entire product.</p><p>The answer was already sitting in our infrastructure diagram.</p><p><strong>Redis executes commands on a single thread.</strong> Not “usually.” Not “with the right configuration.” It is the core of its execution model.</p><p>Commands from every client, across every application server, land in one queue and execute strictly one at a time, to completion. No interleaving. No partial application. No visible intermediate state.</p><p>That is precisely the mutual exclusion we needed, and it costs one network hop.</p><p>An AWS Redis instance handles around 10,000 operations per second, far more than this layer ever asks of it, thus Redis can easily handle these write and read operations.</p><p>Concretely: INCR is atomic. When a hundred Node instances fire INCR activeRequests simultaneously, Redis serialises them. One gets 1, one gets 2, one gets 3.</p><p>Each caller receives back <em>the value after its own increment</em> — exactly the number the naive version failed to compute. The check and the increment collapse into a single operation, and the race disappears.</p><p>For anything more complex than a single command, Redis runs Lua scripts atomically too. The entire script is one uninterruptible unit.</p><p>This is the thing I’d want you to take away from this article.</p><blockquote><em>Redis is usually introduced as “a fast cache.” In a distributed Node.js system, the more valuable property is often its serialisation guarantee, not its speed.</em></blockquote><p>It is a single-threaded coordination point that all of your asynchronous, horizontally-scaled, racing processes can agree on. We weren’t using Redis as a database here. We were using it as a mutex we didn’t have to build.</p><h3>Acquiring and releasing a slot</h3><p>Every call to the ERP begins with an awaited write to Redis. The request increments the counter for each rule that applies to it, and the same round trip returns that counter’s value <em>after</em> its own increment. That is what makes the check safe: there is no read-then-write, so no window for anything else to slip through.</p><p>If the value is within the limit, the request proceeds. If the limit has already been crossed, the request is rejected with a 429 and the increment is rolled back immediately — a request that never ran must not hold a slot. The same applies across rules: if a later rule rejects, every increment already taken is reversed first, otherwise a blocked request would quietly occupy slots it never used.</p><p>The slot is released when the request finishes, on every exit path — success, error and timeout alike. A leaked slot is one the system never gets back.</p><h3>Measuring health in minute buckets</h3><p>The health signals describe the ERP rather than us. Every completed call is recorded against a bucket for the current minute: how long it took, and whether it timed out. The buckets live in Redis, so every application instance contributes to one shared record instead of forming its own private opinion.</p><p>Before a request is sent, the last five minute-buckets are read and aggregated — the share of calls that timed out, and the 90th-percentile response time. If either crosses its threshold the ERP is marked slow and the request is refused. Reading a fixed number of recent buckets keeps the window rolling without scanning unbounded history, and old buckets expire on their own.</p><h3>What the library provides</h3><p>None of the atomic-counter machinery had to be written by hand. rate-limiter-flexible implements it on Redis, with the expiry semantics and production hardening that would have been expensive to reproduce.</p><p>Its model is points rather than requests: a key holds a budget, each request consumes one, and the balance comes back in the same call. Normally that budget refills when a time window rolls over. We invert it.</p><blockquote><em>Our budget refills when work finishes, not when time passes.</em></blockquote><p>A point is taken when a request starts and returned when it completes, which turns the limiter into a semaphore — points become concurrent slots rather than requests per minute, and the limit tightens on its own as the ERP slows. Points can be weighted so an expensive query costs more than a cheap one, and each rule gets its own key prefix so the counters never collide.</p><p>One deliberate exception: if Redis is unreachable or a rule throws, the layer logs and lets the request through, because a limiter that fails closed would cause the very outage it exists to prevent.</p><blockquote><em>Protection systems should degrade toward doing nothing, never toward blocking everything.</em></blockquote><h3>Concluding remarks</h3><p>Two decisions did most of the work here, and neither was about rate limiting as such: what to measure, and where to enforce it. Arrival rate would have been the conventional choice, but because the ERP holds one work process per request for the duration of the call, the quantity that matters is occupancy — and occupancy only tracks arrival rate while service time is stable, which is precisely the assumption that fails as the backend degrades. Counting in-flight requests measures the constraint directly, with latency and timeout signals covering the case where the ERP is struggling for reasons unrelated to us. Enforcing that count across horizontally scaled, asynchronous processes is a shared-mutable-state problem, and Redis’s single-threaded command loop already supplied the serialisation it needed — recognised rather than built, which is probably the most transferable result here. The rest was discipline about not leaking state: release slots on every exit path, roll back partial reservations, refund rejected attempts, expire orphaned counters, and fail open, because a protection layer that fails closed turns a dependency blip into a total outage. rate-limiter-flexible provided the atomic counter; the judgement lay in choosing the unit, defining the scopes, and placing the guard at the one chokepoint every call must pass through, since enforcement that depends on a developer remembering it is a convention rather than a control. The result is a system that declines work it cannot safely perform, quickly and with a reason attached, before any load reaches the ERP.</p><blockquote><em>Sometimes the best possible response to a request is a fast, well-explained no.</em></blockquote><p><em>If you have solved a comparable problem — or met the same race in a different runtime — I would be interested to hear how you approached it.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d14c34f036d2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codeelevation/protecting-a-system-that-cant-say-no-d14c34f036d2">Protecting a System That Can’t Say No</a> was originally published in <a href="https://medium.com/codeelevation">CodeElevation</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[“Typing Code Is Dead!” 18-Year Microsoft Veteran Sparks Backlash: No Wonder Windows 11 Is So Bad]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/typing-code-is-dead-18-year-microsoft-veteran-sparks-backlash-no-wonder-windows-11-is-so-bad-e5f2067fcb4a?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/1080/0*fjS5icJDEDkcmBwM" width="1080"></a></p><p class="medium-feed-snippet">&#x201C;The era of typing code is absolutely over.&#x201D;</p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/typing-code-is-dead-18-year-microsoft-veteran-sparks-backlash-no-wonder-windows-11-is-so-bad-e5f2067fcb4a?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/typing-code-is-dead-18-year-microsoft-veteran-sparks-backlash-no-wonder-windows-11-is-so-bad-e5f2067fcb4a?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/e5f2067fcb4a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[windows]]></category>
            <dc:creator><![CDATA[Dylan Cooper ]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 18:24:23 GMT</pubDate>
            <atom:updated>2026-11-16T18:24:21.983Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[A Programming Language That Took 10 Years to Build Was Made Obsolete by AI in Just 11 Days!]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/a-programming-language-that-took-10-years-to-build-was-made-obsolete-by-ai-in-just-11-days-6948f6b3ec52?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/1080/0*-j44D9oQp7zXLg4s" width="1080"></a></p><p class="medium-feed-snippet">In 2016, a programming language called Zig was born.</p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/a-programming-language-that-took-10-years-to-build-was-made-obsolete-by-ai-in-just-11-days-6948f6b3ec52?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/a-programming-language-that-took-10-years-to-build-was-made-obsolete-by-ai-in-just-11-days-6948f6b3ec52?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/6948f6b3ec52</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[programming-languages]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Dylan Cooper ]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 18:24:21 GMT</pubDate>
            <atom:updated>2026-11-16T18:24:20.254Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Software Runs Slowly Because of Poor Design, Not Because the Programming Language Itself Is Slow]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/software-runs-slowly-because-of-poor-design-not-because-the-programming-language-itself-is-slow-51a6b939dc07?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/1080/0*sya0wA7xIL6QC-FW" width="1080"></a></p><p class="medium-feed-snippet">Recently, quite a few software projects have been rewritten from their original languages into Rust using AI. But is this actually the&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/software-runs-slowly-because-of-poor-design-not-because-the-programming-language-itself-is-slow-51a6b939dc07?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/software-runs-slowly-because-of-poor-design-not-because-the-programming-language-itself-is-slow-51a6b939dc07?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/51a6b939dc07</guid>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[programming-languages]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Deo Wetton]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 18:22:11 GMT</pubDate>
            <atom:updated>2026-11-16T18:22:10.477Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[How I Built a Real-World CI/CD Pipeline for a .NET 10 Application on Azure]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/how-i-built-a-real-world-ci-cd-pipeline-for-a-net-10-application-on-azure-763034d7239a?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/1024/1*ENL0FLO9cJBp8cpmmUfldQ.jpeg" width="1024"></a></p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/how-i-built-a-real-world-ci-cd-pipeline-for-a-net-10-application-on-azure-763034d7239a?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/how-i-built-a-real-world-ci-cd-pipeline-for-a-net-10-application-on-azure-763034d7239a?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/763034d7239a</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[csharp]]></category>
            <category><![CDATA[dotnet]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[  Mori]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 18:10:47 GMT</pubDate>
            <atom:updated>2026-11-16T18:10:45.861Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Why I Use the Unit of Work Pattern With EF Core | Clean Architecture (.NET 10)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/why-i-use-the-unit-of-work-pattern-with-ef-core-clean-architecture-net-10-0970de59e988?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/1024/1*rI440IPGljIV6Zt0LnL6oA.jpeg" width="1024"></a></p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/why-i-use-the-unit-of-work-pattern-with-ef-core-clean-architecture-net-10-0970de59e988?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/why-i-use-the-unit-of-work-pattern-with-ef-core-clean-architecture-net-10-0970de59e988?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/0970de59e988</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[csharp]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[dotnet]]></category>
            <dc:creator><![CDATA[  Mori]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 18:09:13 GMT</pubDate>
            <atom:updated>2026-11-16T18:09:12.662Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Everyone Codes With AI Now. Almost Nobody Trusts It. Here’s the Skill That Fills the Gap.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/everyone-codes-with-ai-now-almost-nobody-trusts-it-heres-the-skill-that-fills-the-gap-5eab1e75999b?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/1024/1*AgXkEc9PFKU8rqTEZ3Grzg@2x.jpeg" width="1024"></a></p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/everyone-codes-with-ai-now-almost-nobody-trusts-it-heres-the-skill-that-fills-the-gap-5eab1e75999b?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/everyone-codes-with-ai-now-almost-nobody-trusts-it-heres-the-skill-that-fills-the-gap-5eab1e75999b?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/5eab1e75999b</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Behind The Code]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 18:08:59 GMT</pubDate>
            <atom:updated>2026-11-16T18:08:58.494Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[7 Clean Code Rules That Backfired On Me]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/7-clean-code-rules-that-backfired-on-me-f7f0688a0f8b?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/1672/1*idJsLIIKKkKzBeC9HAn0YA.png" width="1672"></a></p><p class="medium-feed-snippet">DRY, tiny functions and extra layers looked perfect in code review. Then the next change arrived</p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/7-clean-code-rules-that-backfired-on-me-f7f0688a0f8b?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/7-clean-code-rules-that-backfired-on-me-f7f0688a0f8b?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/f7f0688a0f8b</guid>
            <category><![CDATA[clean-code]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Codebook Fusion]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 18:08:24 GMT</pubDate>
            <atom:updated>2026-11-16T18:58:42.715Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI Coding Success Rate Leaderboard: Go and Rust Crush C++]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/codeelevation/ai-coding-success-rate-leaderboard-go-and-rust-crush-c-2372be3c5981?source=rss----1c377a233996---4"><img src="https://cdn-images-1.medium.com/max/2600/1*kaIhZrEG60cn5Guf7Eck8g.png" width="3222"></a></p><p class="medium-feed-snippet">Over the past two years, the programming community has experienced an unprecedented &#x201C;professional identity crisis.&#x201D;</p><p class="medium-feed-link"><a href="https://medium.com/codeelevation/ai-coding-success-rate-leaderboard-go-and-rust-crush-c-2372be3c5981?source=rss----1c377a233996---4">Continue reading on CodeElevation »</a></p></div>]]></description>
            <link>https://medium.com/codeelevation/ai-coding-success-rate-leaderboard-go-and-rust-crush-c-2372be3c5981?source=rss----1c377a233996---4</link>
            <guid isPermaLink="false">https://medium.com/p/2372be3c5981</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Dylan Cooper ]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 16:42:19 GMT</pubDate>
            <atom:updated>2026-11-11T16:42:18.509Z</atom:updated>
        </item>
    </channel>
</rss>