<?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 Sergii Koval on Medium]]></title>
        <description><![CDATA[Stories by Sergii Koval on Medium]]></description>
        <link>https://medium.com/@kovallux?source=rss-266b9d423608------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*W4dPaozzjicnLrOdPZrwNA.jpeg</url>
            <title>Stories by Sergii Koval on Medium</title>
            <link>https://medium.com/@kovallux?source=rss-266b9d423608------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 09:42:21 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kovallux/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[The deadline nobody circled: in September, your app goes on a 24-hour clock.]]></title>
            <link>https://medium.com/@kovallux/the-deadline-nobody-circled-in-september-your-app-goes-on-a-24-hour-clock-6c9951759b2a?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/6c9951759b2a</guid>
            <category><![CDATA[mobile-apps]]></category>
            <category><![CDATA[regulatory-compliance]]></category>
            <category><![CDATA[medical-devices]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Fri, 10 Jul 2026 12:27:59 GMT</pubDate>
            <atom:updated>2026-07-10T12:27:59.849Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*omu25XD5Lgi6rleceZ9MCQ.png" /></figure><p><em>From this September, the moment a flaw in your app is exploited, you have 24 hours to report it. It applies to apps already shipped, and dodging the “medical device” label does not get you out.</em></p><p>The last four pieces were about what is inside your app and what it costs you. This one is about a date. It falls on 11 September 2026, roughly two months from now, and most of the teams it applies to have not marked it.</p><p>From that day, a new European law starts a clock. If a weakness in your product is being used in a real attack, you have 24 hours to report it to the authorities, with more detail to follow. Not a leisurely disclosure when your PR team is ready. 24 hours from the moment you know.</p><h3>First, whether it catches you</h3><p>The law is the Cyber Resilience Act, and it is deliberately broad. It covers what it calls products with digital elements, which in plain terms is almost any software or connected device sold in Europe. It does not care which industry you are in. And like the rule in the last piece, it reaches past the EU’s borders: if your app is available on the European market, it applies to you from an office outside the EU as much as one inside it. A reader outside Europe who skipped the last piece as “a European thing” should read this one.</p><p>There is a catch that health companies in particular try to use, and it does not work. If your app is a regulated medical device, it sits under the separate medical-device rules instead, which are being rewritten right now to add the very same reporting duty. If your app is not a medical device, a wellness or fitness or meditation app that carefully avoided that label, then the Cyber Resilience Act catches it as an ordinary connected product. There is no third category where a health app reaches European users and owes nothing. And if the device itself is exempt, the companion app and the cloud around it usually are not. Dodging the medical-device label does not buy you out of the rules. It just changes which set you answer to.</p><h3>What actually has to be reported</h3><p>Now, before anyone panics: this is narrower than it sounds. You do not have to report every bug. The clock only starts for a weakness that is actively being exploited, meaning there is real evidence someone is using it in an attack, or for a serious security incident. Ordinary flaws that you find and fix in the normal way stay out of it. This is the honest part, and it matters, because a piece that tells you to report everything is wrong and you should not trust it.</p><p>But two things make it heavier than it first looks. It applies to products you have already shipped, not just new ones. An app you released in 2019 and half-forgot is in scope the day this starts, if it is still out there and a flaw in it gets exploited. And the fines for getting it wrong reach up to €15 million, or 2.5 per cent of your worldwide turnover, whichever is the bigger number.</p><h3>Why this is the hardest one to fake</h3><p>This deadline is harder to bluff than the others, for a simple reason. The earlier rules reward paperwork. You can write a policy, hold a workshop, keep a folder, and look prepared. A 24-hour clock does not care about your folder. To report an exploited weakness within a day, you first have to notice it, and you can only notice a problem in a part of your product you actually know about.</p><p>That is where four pieces of this series come together. If you do not know what is inside your own app, which secrets are in it, which components it is built from, which of them a stranger could already be using, then you cannot start the clock on time. You will spend the 24 hours discovering what you shipped, and the deadline will pass while you are still reading your own settings files. The plaintext database, the hardcoded key, the forgotten staging address: those are the things that get exploited, and they are the same things you cannot report on quickly if you never looked inside the file to begin with.</p><h3>One last question</h3><p>This is the last of the five pieces, and the question that holds them together is short. You have a compliance file that says you are fine. You have an app that a stranger can open like a folder. On the day a flaw in it is exploited, one of those two things will decide how your next 24 hours go. Ask the people who build your product a single question, and make them answer it plainly: do we know what is inside the app we ship, well enough to defend it, report it in a day, and put a name to it?</p><p>If the answer is yes, and they can show you, you are in a small minority and you should be glad of it. If the answer is a pause, you have just found the most valuable thing in this whole series: the gap, while it is still yours to close, and not yet a headline with your name in it.</p><p><em>This was the last of five short pieces on what sits inside medical apps and which rules they break. If you have read all five, you already know more than most of the people shipping these apps.</em></p><p><em>Sèrge Koval is a security architect in Luxembourg, working on iOS and macOS. He wrote</em> Secure Development of iOS and macOS Apps <em>and takes apart real App Store apps for a living. Finding these things before someone else does is the work he does. He is on LinkedIn at linkedin.com/in/skoval.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6c9951759b2a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[In Europe, a breach is now the executive’s problem, with their name on it.]]></title>
            <link>https://medium.com/@kovallux/in-europe-a-breach-is-now-the-executives-problem-with-their-name-on-it-8c885d69fe11?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c885d69fe11</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[nis2]]></category>
            <category><![CDATA[medical-devices]]></category>
            <category><![CDATA[apps]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Mon, 29 Jun 2026 17:20:50 GMT</pubDate>
            <atom:updated>2026-06-29T17:20:50.949Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-qOay_B_2Bm8nLZqgZv__g.png" /></figure><p><em>NIS2 moves cybersecurity onto the board, makes it personal, and turns a hardcoded password into a question about whether you did your job.</em></p><p>The last piece ended with a fine: large, paid by the company, survivable for a big one. This piece is about a European rule that changes who pays, and it is the part that tends to wake executives up. The money still comes out of the company. The liability now has a person’s name on it.</p><p>First, whether this is even your problem, because the rule is European, but its reach is not. NIS2 mainly catches medium and larger organisations, roughly fifty staff or ten million euros in turnover and above. If your company operates in the European Union, sells into a regulated sector there, or supplies someone who does, it can apply to you even from an office in another country. A health business based abroad with an EU branch is caught the moment that branch is involved. So if you are a funded health company with real users, assume you qualify and check. If you are two people in a garage, you have a little more room, for now.</p><p>The rule is called NIS2, and the detail that matters is simple. It takes cybersecurity, which used to live with the IT team or the security person, and puts it on the management body: the board, the executives, the people who sign decisions. Under NIS2 they must approve the company’s security measures and oversee that they actually happen. Not delegate and forget. Approve, in writing, and supervise.</p><h3>The part with your name on it</h3><p>In practice, “personal” means several things, and none of them are comfortable.</p><p>If your company fails to put adequate security in place and something goes wrong, the consequences do not stop at the company’s bank account. The company can be fined up to €10 million, or two per cent of its worldwide turnover, for the most serious cases, and somewhat less for lower-tier ones. That part is familiar. The rest is new.</p><p>The executive responsible can be held personally liable for the failure. In several countries, that means a personal fine, separate from the company’s. The authority can name that person in public, so the failure is attached to them and not to a faceless brand. And for the most critical operators, an authority can ask a court to ban the responsible executive from holding any management position, and not only at the company where it went wrong. A temporary professional disqualification for getting security wrong.</p><p>One more thing makes this sharper than ordinary corporate liability. It does not require you to have been reckless. The rule allows liability for the failure itself, even when nobody was negligent in the dramatic sense. And in Germany, which has gone furthest, a company is not even allowed to quietly waive its claims against a director for this. Your own board cannot let you off.</p><h3>This is not a future problem</h3><p>It would be comforting to file this under “something to deal with later.” That option has closed. NIS2 has been law across the Union since late 2024. Most member states have now written it into their national law, and the European Commission has hauled several of the slower ones to court for dragging their feet. 2026 is the year the audits and the fines begin in earnest, and the first public penalties have already landed.</p><h3>Your developer’s shortcut, your signature</h3><p>Now connect it to the three pieces before this one. A hardcoded database password, a plain file of patient records on every phone, a live key sitting in the bundle: in the earlier pieces, these were technical findings. Under NIS2 they become something else. They are evidence about whether the management body did its job.</p><p>When an authority asks whether you put adequate security measures in place, “our developer left the database password in a settings file” stops being a story about one developer and becomes a question about whether anyone with a signature was overseeing security at all. The thing your engineer did in a hurry, two years ago, becomes the thing that answers a question about you. That is the shift NIS2 makes: it turns a quiet technical mistake into a line in your personal file.</p><h3>One question for the person who signs</h3><p>If you are the founder, the chief executive, or the person who would be named, the question is for you. If you are not, it is the question to put in front of them, before a regulator does: can we show, on paper, that someone at the top approved our security measures and checked that they are real? The proof an authority wants is a named person who reviewed the measures, signed them, and could survive being questioned about them, not a policy gathering dust in a folder.</p><p>If the honest answer is a shuffle of paper and a hope, you have found the gap while it is still cheap to fix. The alternative is finding it in a letter that has your name on it.</p><p><em>This is the fourth of five short pieces on what sits inside medical apps and which rules they break. Next: the deadline most of you have not noticed, the one that starts this September.</em></p><p><em>Sèrge Koval is a security architect in Luxembourg, working on iOS and macOS. He wrote</em> Secure Development of iOS and macOS Apps <em>and takes apart real App Store apps for a living. He is on LinkedIn at linkedin.com/in/skoval.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8c885d69fe11" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Health data is the expensive kind to lose. The worst part is the leak you can’t measure.]]></title>
            <link>https://medium.com/@kovallux/health-data-is-the-expensive-kind-to-lose-the-worst-part-is-the-leak-you-cant-measure-d6be0977e52f?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/d6be0977e52f</guid>
            <category><![CDATA[cyber-security-awareness]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[hacker]]></category>
            <category><![CDATA[medical-devices]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Tue, 16 Jun 2026 19:24:41 GMT</pubDate>
            <atom:updated>2026-06-16T19:24:41.949Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v31100urN2mC8cy5qWN06w.png" /></figure><p><em>The same bug costs more in a health app, and “we don’t know who was affected” is the answer that turns an incident into a headline.</em></p><p>Your app is a zip file anyone can open, and last time I listed the four things I find inside. Suppose one of them is real: a stranger pulls a plain database of patient records off a phone, or finds a live key that opens your back end. The next question is what it costs you, and the answer turns on one fact. This is health data.</p><h3>Health data is its own category</h3><p>Not all personal data is treated the same. Your customers’ names and email addresses are worth protecting, and losing them is bad. Losing the fact that someone uses a therapy app, a fertility tracker, or an HIV-status tool is worse, because that fact can cost someone a job, or out a private diagnosis to people they never told. Regulators understand this, so they put health data in a heavier class of its own.</p><p>In Europe, the law calls this a special category and reserves its heaviest penalties for mishandling it: up to €20 million, or four per cent of a company’s worldwide turnover, whichever is larger. In the United States the dedicated health-privacy law, HIPAA, covers hospitals, insurers, and the companies that work for them. The label and the numbers differ across countries. The message does not: this is the data you least want to lose.</p><h3>The trap: “we are not a medical device, so this does not apply”</h3><p>A lot of health apps assume the strict rules are for someone else. They are a wellness app, not a medical device. They are not a hospital, so HIPAA is not their problem. For a growing number of them, that assumption is now wrong, and it costs them.</p><p>In the United States, a federal rule updated in 2024 covers exactly these apps: the health and wellness products that sit outside HIPAA. Under it, a “breach” is not only a hacker breaking in. Quietly sharing a user’s health data with the advertising and analytics code baked into your app counts too, and the regulator has already fined well-known apps for doing it. On top of that, new state laws go further. Washington’s health-data law lets the affected people sue you directly, with no need to wait for a regulator and no minimum company size to hide behind.</p><p>In Europe, the special-category rules apply to health data whether or not you ever call your app a medical device. Choosing the lighter regulatory label for your product does not move your users’ data into a lighter class. The data is what it is, and the law follows the data.</p><h3>The clock starts when you find out</h3><p>When health data leaks, you do not get to investigate quietly for a few months and announce it when you are ready. Europe expects you to report a serious breach to the regulator within about three days of discovering it. The American rules and the state laws have their own deadlines, and some now let your customers take you to court while you are still working out what happened. The moment you know, the clock is running, and “we were still looking into it” is not a defence anyone accepts.</p><h3>The leak you cannot measure</h3><p>All of this ties back to the file on the phone, in a way that ought to worry you more than the fine.</p><p>To report a breach honestly, you have to answer a simple question: whose data, and how much? If your patient records lived in a single system on your servers, with logs of who touched them, you can answer it. You can say which accounts were exposed and when, and notify those people. The incident is contained because it is measurable.</p><p>A plain database shipped inside the app is the opposite. Once that file is on a customer’s phone, you have no record of anything. You cannot tell who opened it, who copied it, or whether it was read once or a thousand times. If it leaks, you cannot scope the damage, because the damage happened on devices you never see. The only honest breach notice you can write is the one no board wants to sign: we believe every user may be affected, and we cannot narrow it down.</p><p>That is the real cost of the findings from the last two pieces. The encryption on your server, the one your security page is proud of, protects the copy of the data you can already see. It does nothing for the copy you handed to every person who installed the app. A leak you can measure is a bad week. A leak you cannot measure is the one that ends up with your company’s name in it.</p><h3>One question to take upstairs</h3><p>Ask the people who run your app one thing: if we had to file a breach notice tomorrow, could we say exactly whose health data was exposed, and how much? A confident answer means your data lives somewhere you can see and measure. A pause, or a shrug, means it does not, and that the breach notice would be the bad kind. Either way you have learned something, on a quiet afternoon, instead of in a letter from a regulator.</p><p><em>This is the third of five short pieces on what sits inside medical apps and which rules they break. Next: the European rule that stops treating this as your company’s problem and starts naming you, personally.</em></p><p><em>Sèrge Koval is a security architect in Luxembourg, working on iOS and macOS. He wrote</em> Secure Development of iOS and macOS Apps <em>and takes apart real App Store apps for a living. He is on LinkedIn at </em><a href="https://www.linkedin.com/in/skoval/"><em>linkedin.com/in/skoval</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d6be0977e52f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Four things I find inside medical apps, and how to tell the dangerous from the harmless]]></title>
            <link>https://medium.com/@kovallux/four-things-i-find-inside-medical-apps-and-how-to-tell-the-dangerous-from-the-harmless-98d205f4b553?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/98d205f4b553</guid>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[medtech]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[medical-devices]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Fri, 12 Jun 2026 04:29:15 GMT</pubDate>
            <atom:updated>2026-06-12T04:29:15.764Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ccsmr-NQRDyhiYkVBt84nA.png" /></figure><p><em>Not everything that looks alarming is dangerous. The difference is the whole job.</em></p><p>Your app is a zip file anyone can open. No audit you passed involved opening it. That was the last piece; this one is the inventory.</p><p>I sort what I find by a single test: could a stranger use this against you? For a fair amount of it, no, and saying so out loud is most of what people pay me for.</p><h3>1. The database that opens without a key</h3><p>Most health apps keep a copy of your data on the phone, in a file called a database. That part is normal, because apps need to work offline, so they cache things locally. The question is what is in that file and whether anyone can read it. In app after app, the answer is patient records, session notes, moods, sleep, identifiers, sitting in plain text. No lock. A free database browser opens it like a spreadsheet.</p><p>It gets worse when the key is in the box with the lock. More than once I have found the database password written into a settings file shipped inside the same app, in plain text. The password that should protect the data sits a folder away from the data it protects. Whoever wrote it assumed nobody would look. People look.</p><p>This is the finding that should worry you most. No clever attack, no breaking in: the data is already readable on every phone that installed the app.</p><h3>2. Secrets baked into the app, real and fake</h3><p>Open the settings files and you find keys and tokens, the short strings that let your app talk to other services for payments, analytics, login and messaging. In bulk they look alarming. Most of them are not a problem.</p><p>A lot of these keys are meant to live in the app and be seen. They are public by design, like the address printed on your shopfront. A payment provider gives you one key for the customer’s phone and keeps the real one on its own servers. Finding the phone key proves nothing.</p><p>A few are different. A live production key, a server password, a token that was never meant to leave your building. Those turn up too, in the same file as the harmless ones, with nothing to mark them apart. One real key can let a stranger run up charges in your name, or walk in through a door you forgot was open. Finding keys is easy. Knowing which of the forty you found is the one that matters is the job.</p><h3>3. A free map of your back end</h3><p>Some of what I find is not a single secret but a layout. The internal web addresses your team uses for testing, the ones named things like api-staging. The details of your cloud project. The identifiers your login system uses across your apps. No single item is a password. Together they are a blueprint of how your system is built and where the soft parts are.</p><p>A blueprint matters even if you never write a line of code, because attackers are lazy in a useful way. They try the easy way in before the hard one. Your test systems are usually softer than your live ones, built fast and watched less. Shipping their addresses to every customer hands a stranger a to-do list with the weak items already marked.</p><h3>4. The things that should never be in there</h3><p>Then there is the category I cannot predict, where a developer has done something strange. The credentials that prove your servers are yours, left in the bundle, so a stranger could impersonate them. The private logic of the product, the actual rules behind its health predictions, written into a plain text file the app reads and runs, so anyone can read the recipe and copy it. An encryption key for users’ private messages, sitting in the open, which means the messages are not private at all.</p><p>These are rare, and they are the ones that end up in a news story. Someone made each of them on purpose, without thinking it through, and they are specific enough that a regulator or a competitor knows exactly what they are reading.</p><h3>What is not a leak</h3><p>Most people selling fear leave this part out, so I will say it straight: a long list of findings is not the same as a long list of problems.</p><p>A scan of a normal app produces a frightening report with dozens of items flagged. Open them one by one and most are the public-by-design keys from earlier, doing their job: a test payment key, the analytics identifiers every app on your phone carries, a cloud config file the provider itself says is safe to ship. Calling those a data leak in a headline is how you lose the one engineer in the room who knows better, and that engineer is usually the person your boss turns to.</p><p>An honest report does two jobs. It names the small number of things that are dangerous, and it tells you with equal clarity what you can ignore. A report that only does the first half is selling you a fright.</p><h3>What to do with this</h3><p>You now have four categories to ask about. Take them to whoever builds your app, in plain words, and watch which one makes them shift in their seat. Is any patient data stored unencrypted on the phone, and is the password to it sitting in the app as well? Are there live keys or server passwords in the settings files? Do you ship the addresses of your internal or test systems? Is there anything in the bundle that runs code, or holds a key, that you would not want read aloud in a meeting?</p><p>A confident team answers all four in a couple of sentences, because they already checked. A team that goes quiet on one of them has shown you where to look first. Either way you learn something, and you learn it before someone less friendly does.</p><p><em>This is the second of five short pieces on what sits inside medical apps and which rules they break. Next: why health data counts as a special category, and why a leak you cannot measure is the worst kind to report.</em></p><p><em>Sèrge Koval is a security architect in Luxembourg, working on iOS and macOS. He wrote</em> Secure Development of iOS and macOS Apps <em>and takes apart real App Store apps for a living. He is on LinkedIn at </em><a href="https://www.linkedin.com/in/skoval/"><em>linkedin.com/in/skoval</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98d205f4b553" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[That Box of Old Family Videos? Your iPhone Can Fix Them Now.]]></title>
            <link>https://medium.com/@kovallux/that-box-of-old-family-videos-your-iphone-can-fix-them-now-90557b1adfd4?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/90557b1adfd4</guid>
            <category><![CDATA[repair-video]]></category>
            <category><![CDATA[family-video]]></category>
            <category><![CDATA[video-upscaler]]></category>
            <category><![CDATA[video-enhancer]]></category>
            <category><![CDATA[old-videos]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Wed, 10 Jun 2026 05:52:04 GMT</pubDate>
            <atom:updated>2026-06-10T05:52:04.274Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r0ZtO6W06qwPLZn0muLEDg.png" /></figure><p><em>AI video upscaling on your phone is real, it works, and you don’t need a computer or any technical knowledge. Here’s how to rescue your family’s video memories in 2026.</em></p><h3>You probably have footage that deserves better</h3><p>Somewhere in your life, there’s video footage that matters to you but looks terrible.</p><p>Maybe it’s your parents’ wedding recorded on a camcorder in the 90s. Your kid’s first steps captured on an iPhone 5. A holiday video from 2012 that Facebook compressed into a pixelated mess. A graduation ceremony filmed from the back row. Your grandparents on a VHS tape that someone eventually digitised into a blurry 480p file.</p><p>This footage is irreplaceable. And it looks like it was filmed through a dirty window.</p><p>For years, the only option to improve it was expensive desktop software — tools like Topaz Video AI, which cost real money, require a powerful Windows or Mac computer, and have a learning curve that assumes you know what a “codec” is.</p><p>That’s changed. Modern iPhones have dedicated AI hardware — Apple’s Neural Engine and GPU — that can run genuine upscaling models directly on the device. No computer. No uploads to mysterious cloud services. No technical expertise. The same kind of neural networks that professional tools use, running right on the phone in your pocket.</p><h3>What “upscaling” actually does (in plain language)</h3><p>When you record video at low resolution — say 480p or 720p — the image contains a limited number of pixels. Each frame is a grid of coloured dots, and there simply aren’t enough dots to show fine detail.</p><p>AI upscaling adds more pixels. But not randomly — the models have been trained on millions of video frames to <em>predict</em> what the missing detail should look like. They’ve learned that this particular blur pattern usually means a human face, that this texture is probably grass, that this edge should be sharper.</p><p>The result isn’t magic. You can’t turn a 1990s camcorder recording into footage that looks like it was shot yesterday on an iPhone 16. But you can make it look <em>significantly</em> clearer, sharper, and more watchable on a modern screen.</p><p>Think of it as going from “I can barely tell who’s in this video” to “Oh, that’s clearly Uncle Pete at the barbecue.”</p><h3>What you need</h3><ul><li>An iPhone running iOS 17 or newer (also works on iPad, Mac with Apple Silicon, and even Apple Vision Pro)</li><li>The video file on your phone (in your Photos library, Files app, or transferred via AirDrop)</li><li>A video upscaling app</li><li>A few minutes of patience while it processes</li></ul><p>That’s it. No computer. No cables. No accounts to create. No footage uploaded anywhere.</p><h3>How it works in practice</h3><p>I’ll walk through the actual process using my app as an example, since I built it specifically for this.</p><ol><li><strong>Open the app and select a video</strong> from your library</li><li><strong>Choose your AI model:</strong></li></ol><ul><li><strong>Apple SuperRes or MetalFX</strong> (free, fast) — good for footage that’s already decent but needs more resolution. These use Apple’s own frameworks, including technology borrowed from AAA gaming, to intelligently predict missing pixels.</li><li><strong>BSRGAN</strong> (premium) — the best option for seriously degraded footage. Old VHS transfers, heavily compressed downloads, very grainy clips. It’s slower but recovers remarkable detail.</li><li><strong>Real-ESRGAN</strong> (premium) — produces the most photorealistic-looking output. Best for footage where you want the result to look natural rather than “processed.”</li></ul><ol><li><strong>Pick your target resolution</strong> — 4K is usually the sweet spot for old footage. The free tier goes up to 4K. Premium unlocks 8K and 16K for future-proofing or large-format displays.</li><li><strong>Hit convert</strong> and wait. Processing time depends on the video length, your device, and the model you chose. The fast models (SuperRes/MetalFX) process in near real-time. The premium models take longer but produce better results.</li><li><strong>Use the comparison view</strong> to see before and after side-by-side. Zoom in to see the detail difference clearly — on a phone screen at normal zoom, the difference can be subtle. It becomes obvious when you zoom in or watch on a larger screen.</li><li><strong>Save or share</strong> the upscaled version.</li></ol><p>Everything happens on your device. No internet connection needed. You could do this on a plane.</p><h3>What to expect: honest results</h3><p>I want to be straight about this, because a lot of apps overpromise.</p><p><strong>What works well:</strong> — Old iPhone footage (pre-iPhone 8 era) → noticeably sharper and cleaner — Social media downloads that got compressed (Facebook, WhatsApp, Instagram saves) → restored to near-original quality — Webcam recordings → significantly improved — Digitised VHS/camcorder footage → improved clarity, especially with BSRGAN, though it’ll still look “vintage” — Dark or grainy footage → reduced noise, better detail in shadows</p><p><strong>What doesn’t work miracles:</strong> — Extremely low resolution (240p or below) → improvement is visible but limited. The app will warn you when the resolution jump is too extreme. — Severely motion-blurred footage → upscaling sharpens static detail, not motion blur — Audio quality → this is video upscaling only. Bad audio stays bad.</p><p><strong>Processing time (rough guide):</strong> — A 30-second clip with MetalFX/SuperRes: under a minute — A 30-second clip with BSRGAN: a few minutes — Longer videos scale proportionally</p><p>The phone will get warm during processing. This is normal — it’s running AI computations on dedicated hardware. Don’t worry about it.</p><h3>The privacy thing most people overlook</h3><p>Here’s something worth knowing: most online video “enhancers” you find via Google require you to upload your video to their servers. Your family footage — your kids, your home, your personal moments — goes to a server farm owned by a company you’ve never heard of, gets processed, and comes back.</p><p>Is this likely to cause a problem? Probably not. But it’s also completely unnecessary.</p><p>On-device upscaling means the AI runs on your phone’s chip. The video never leaves your device. No internet connection required — you can do it in airplane mode. For family videos especially, this just feels right. These are your memories. They should stay on your hardware.</p><h3>Which videos to start with</h3><p>If you’re going to try this, start with footage that has sentimental value but isn’t too long. A 30-second clip of your kid at age 3. A short segment from a family gathering. A holiday moment.</p><p>The satisfaction of seeing a blurry memory become clear is surprisingly emotional. Several people I’ve talked to have described it as “getting that moment back.” That might sound dramatic until you try it with your own footage.</p><p>Once you’ve done a short clip and you like the results, work through your library. Old footage from Facebook or WhatsApp downloads is particularly rewarding to upscale because those platforms compress video aggressively — the original quality was almost certainly much better than what you saved.</p><h3>A note on “16K” and resolution marketing</h3><p>You’ll see apps (mine included) advertising 8K or 16K output. Let me be honest: for old family footage, you almost never need anything above 4K. Your TV is probably 4K. Your phone screen is effectively smaller than 4K. Going beyond that has technical uses (future-proofing archival copies, professional workflows, large-format displays) but for family videos, 4K is the practical ceiling.</p><p>The free tier covers this. You get unlimited conversions at up to 4K using Apple SuperRes and MetalFX. That’s enough for the vast majority of family video restoration.</p><p>Don’t get caught up in resolution numbers. The quality of the upscaling — how well the AI reconstructs detail — matters far more than how large the output resolution is.</p><h3>The best time to do this is now</h3><p>AI upscaling models improve every year, so in theory you could wait. But footage degrades over time — files get corrupted, storage fails, formats become obsolete. The people in those videos get older. Some won’t be around forever.</p><p>The technology to rescue these memories exists today, runs on the phone in your pocket, costs nothing to try, and keeps your footage completely private. There’s no good reason to leave those videos sitting in a drawer — digital or physical — looking worse than they need to.</p><p>Your family filmed those moments because they mattered. Give them the quality they deserve.</p><p><em>I’m Sergii Koval, the developer behind </em><a href="https://apps.apple.com/app/id1494184431"><em>Video Upscaler 16K</em></a><em>. It runs on iPhone, iPad, Mac (Apple Silicon), and Apple Vision Pro. Four AI models, 53 languages, 100% on-device processing. The free tier gives you unlimited conversions up to 4K — enough to see if it works for your footage before spending anything.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*viniQrmXCPqnmrHwqtjkyg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=90557b1adfd4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your medical app passed every audit. Nobody looked inside it.]]></title>
            <link>https://medium.com/@kovallux/your-medical-app-passed-every-audit-nobody-looked-inside-it-f8bf1ebd5ba9?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/f8bf1ebd5ba9</guid>
            <category><![CDATA[mobile-security]]></category>
            <category><![CDATA[digital-health]]></category>
            <category><![CDATA[data-privacy]]></category>
            <category><![CDATA[medical-devices]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Tue, 09 Jun 2026 18:21:30 GMT</pubDate>
            <atom:updated>2026-06-09T18:21:30.271Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SnOL8L8pWt3gRJXD3w1lpA.png" /></figure><p><em>A compliance certificate says nothing about what your app gives away.</em></p><p>You run a health app. You have the paperwork: a signed data agreement, a privacy policy a lawyer approved, perhaps an ISO certificate on the wall. Apple let the app into the store. Patient data travels to your servers through an encrypted connection. By every measure your company tracks, the app is compliant.</p><p>None of that involved anyone opening the actual file your patients download and reading what sits inside it.</p><p>That file is your product. When a patient installs your app, their phone downloads a package called an .ipa. The name sounds technical. It is a zip archive wearing a different extension. Rename it, double click, and it opens like a folder of holiday photos. No password, no jailbroken phone. A teenager with a laptop can do it, and many have.</p><p>Inside that folder is everything your app needs to run on its own: configuration files, images, the compiled code, and the settings your developers fixed in place before shipping. Some of those settings should never leave your office. In the apps I take apart, they often do.</p><h3>What the audits actually check</h3><p>A compliance review reads your documents and your processes. It asks whether you wrote down how you handle patient data, whether your staff signed the right forms, whether you have a plan for a breach. It is a reading exercise about how your company behaves on paper. Useful work, and necessary. It does not put anyone in front of your app looking for secrets left inside it.</p><p>Apple’s review is a separate job. Before your app reaches the store, Apple checks that it obeys the store rules and avoids the private programming shortcuts Apple forbids. The reviewer protects the platform and the user from obvious bad behaviour. The reviewer does not hunt for a database password you left in a text file. That was never the purpose of the review, and Apple has never pretended it was.</p><p>Then there is the encryption everyone mentions first. Your data travels to the server through a protected tunnel, and the server keeps it in encrypted form. Both true, both good. Neither one protects the copy of your app sitting on every customer’s phone. The tunnel guards data while it moves. The file on the device sits still, in the open, on a machine you do not own.</p><p>Each of these checks does its own job well. A gap sits in the middle, where the shipped file lives. Your compliance officer assumes the developers cleaned it. The developers assume the store review would have flagged anything dangerous. Neither of them opened it. The file goes out to every install with whatever is inside.</p><h3>What I find inside</h3><p>When I open these files, the same things keep turning up. A configuration file with a database password written in plain text, readable by anyone. Lists of internal web addresses a company uses for testing, the kind never meant to face the public. Keys and tokens that let one piece of software talk to another, some harmless, some not. In one meditation app, the developer had written part of the app’s private logic into a plain text file that the app reads and runs. The rules behind its health predictions sat there for anyone to read, like a recipe taped to the outside of the safe.</p><p>Most scare stories about this fall apart at the next step, so here is the honest part. A good number of those keys belong in the app. They are meant to be public, the same way your office address is public. Anyone who knows the field can tell which is which in seconds, and an honest report tells you plainly. The danger is the small number of items that were supposed to stay private: the database password, or the internal logic that runs your product. One of those is enough to ruin a quarter.</p><h3>Why this is worse for a medical app</h3><p>For an ordinary app, leaving this lying around is careless. For a medical app it becomes a regulatory matter, because the rules that govern medical software expect security to live in the product, not only in your filing cabinet. In Europe, the Medical Device Regulation says the product must be built to a current security standard, with protection against people who should not have access. In the United States, a hardcoded password guarding patient records is the kind of thing regulators read as a plain security failure. Different wording, same expectation: the file on the phone has to be clean. A journalist or a bored attacker can check it in an afternoon, and the file does not care what your certificate says.</p><p>This is the part that tends to send people back to their teams with questions, which is the right instinct.</p><h3>One question to ask</h3><p>You do not have to take my word for any of this. You need one answer from the people who built your app. Ask them: if someone renames our app file to .zip and opens it, what is inside, and is any of it a secret? If they answer in a sentence, they have checked. If they go quiet, you have found your next problem before a regulator did.</p><p><em>This is the first of five short pieces on what sits inside medical apps and which rules they break. Next: the four things I find most often, and which ones should worry you.</em></p><p><em>Sèrge Koval is a security architect in Luxembourg, working on iOS and macOS. He wrote</em> Secure Development of iOS and macOS Apps <em>and takes apart real App Store apps for a living. He is on LinkedIn at </em><a href="https://www.linkedin.com/in/skoval/"><em>linkedin.com/in/skoval</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f8bf1ebd5ba9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Repurposed Apple’s Gaming Framework to Upscale Video on iOS — Here’s How MetalFX Actually Works…]]></title>
            <link>https://medium.com/@kovallux/i-repurposed-apples-gaming-framework-to-upscale-video-on-ios-here-s-how-metalfx-actually-works-0ce26645fa7c?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/0ce26645fa7c</guid>
            <category><![CDATA[video-enhancer]]></category>
            <category><![CDATA[game-development]]></category>
            <category><![CDATA[video-production]]></category>
            <category><![CDATA[video-upscaler]]></category>
            <category><![CDATA[apple-metal]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Sat, 21 Mar 2026 07:41:09 GMT</pubDate>
            <atom:updated>2026-03-21T07:41:09.959Z</atom:updated>
            <content:encoded><![CDATA[<h3>I Repurposed Apple’s Gaming Framework to Upscale Video on iOS — Here’s How MetalFX Actually Works for This</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5vJdCw0rFdfYyQ-z_nBi7Q.jpeg" /></figure><p><em>MetalFX was designed to make AAA games render faster on Apple Silicon. It turns out the same spatial upscaling technique works remarkably well on pre-recorded video. Here’s what I found building a production app around it.</em></p><h3>The problem: AI video upscaling on iOS was basically impossible</h3><p>Image upscaling on iOS has been doable for a while. Take a CoreML model, feed it a single image, get back a larger image. Straightforward.</p><p>Video is a completely different problem.</p><p>A 30-second video at 30fps is 900 frames. Each frame needs to be extracted, upscaled through an AI model, and reassembled into a video stream with correct timing, audio sync, and codec encoding. On a desktop with an RTX 4090, this is manageable. On an iPhone, running a full BSRGAN or Real-ESRGAN model on 900 frames through CoreML was taking roughly 2–3 seconds per frame on an iPhone 15 Pro in my testing. For a 30-second clip, that’s 30–45 minutes of processing. For a 5-minute video? Hours. Not a product. A punishment.</p><p>The CoreML path works — the models produce excellent results — but the processing time made it impractical as the only option. I needed a fast path alongside the quality path.</p><h3>MetalFX: the gaming tech that accidentally solves video upscaling</h3><p>Apple introduced MetalFX at WWDC 2022 as a performance technology for games. The idea: instead of rendering every game frame at native resolution (expensive), render at lower resolution and let MetalFX intelligently upscale it in real time. It’s conceptually similar to AMD’s FSR — in fact, Apple’s legal disclosures confirm MetalFX is built on AMD’s open-source FidelityFX Super Resolution technology, adapted for Apple’s platforms.</p><p>The key: MetalFX doesn’t just stretch pixels. Its spatial scaler (MTLFXSpatialScaler) analyses the input’s spatial information — edges, textures, patterns — and predicts what the higher-resolution version should look like. It’s running on the GPU, it’s fast, and it’s designed to produce results that look natural on a display.</p><p>The question I had: this thing was built for live game frames being generated by a rendering engine. Could it work on pre-recorded video frames extracted from a file?</p><p>Short answer: yes. And it’s fast enough to be practical.</p><h3>What I actually built</h3><p>The app (Video Upscaler 16K) ships four AI models in a single iOS app, each with different speed/quality trade-offs:</p><p><strong>Free tier (fast path):</strong> — <strong>Apple SuperRes</strong> — Apple’s native super-resolution, optimised for their hardware — <strong>MetalFX</strong> — GPU-accelerated spatial upscaling, repurposed from the gaming framework</p><p><strong>Premium tier (quality path):</strong> — <strong>BSRGAN</strong> — Top-ranked CoreML model for detail recovery and restoration. Excels at heavily degraded footage. — <strong>Real-ESRGAN</strong> — Industry-standard enhancement model. Best for photorealistic output on footage that’s decent quality but needs more resolution.</p><p>The free tier handles conversions up to 4K. Premium unlocks 8K and 16K output, plus the heavier CoreML models, a video editor (trim, crop, effects), and AI metadata management.</p><p>All four models run entirely on-device. No server. No cloud. No internet connection required. The video never leaves the phone.</p><h3>The MetalFX implementation: what actually happens</h3><p>Here’s the pipeline for the MetalFX path:</p><ol><li><strong>Extract frames</strong> from the source video using AVAssetReader</li><li><strong>Convert each frame to a Metal texture</strong> — this is where you bridge AVFoundation’s pixel buffers into Metal’s texture format</li><li><strong>Run MTLFXSpatialScaler</strong> — feed the low-res texture in, get a high-res texture out</li><li><strong>Convert the output texture back to a CVPixelBuffer</strong> for AVAssetWriter</li><li><strong>Encode the upscaled frames</strong> back into a video file with proper timing and codec settings</li><li><strong>Mux the original audio track</strong> back in — upscaling is visual only, audio passes through unchanged</li></ol><p>MetalFX spatial upscaling on an iPhone 15 Pro processes a 720p → 4K frame in roughly 8–15 milliseconds. That’s real-time-capable. For a 30fps video, you need each frame under 33ms, and MetalFX easily beats that threshold.</p><p>By contrast, running BSRGAN through CoreML on the same device: 2000–3000ms per frame. MetalFX is roughly 100–200x faster. The quality difference is real — BSRGAN produces measurably better detail reconstruction on degraded footage — but for a lot of content (recent iPhone footage, social media downloads, anything that’s already decent quality), MetalFX’s output is more than sufficient.</p><h3>Things that went wrong (so you don’t repeat them)</h3><p><strong>Memory pressure.</strong> Processing video frames through Metal means allocating and releasing GPU textures at 30+ per second. If you’re not aggressive about reusing textures, you’ll get memory pressure warnings and eventually crash. A texture pool/ring buffer pattern is essential.</p><p><strong>Thermal throttling.</strong> Sustained GPU workload on an iPhone generates heat. After about 2–3 minutes of continuous upscaling, the chip throttles and per-frame time increases. For longer videos, the progress estimate needs to be dynamic — don’t promise “3 minutes remaining” and then have it take 8. I display a real-time estimate that adjusts as throughput changes.</p><p><strong>The 16K question.</strong> Can you actually output 16K from MetalFX on a phone? Technically yes. Practically, it depends on the source material and the device. There’s a minimum pixel density needed in the input for the output to be meaningful at extreme resolutions. If someone tries to upscale a 240p clip to 16K, the result won’t be what they expect. The app handles this by analysing the input and warning when the resolution jump is too extreme.</p><p><strong>Comparison view was essential.</strong> Early testers couldn’t tell if the upscaling had worked because they were viewing the output on the same phone screen. The input and output look identical at phone-screen size. Adding a side-by-side comparison view where you can zoom into details was critical for users to actually see and appreciate the quality improvement. This is doubly important for conversion — people need to see value before they pay.</p><h3>Why on-device matters (beyond the technical challenge)</h3><p>Every cloud-based video upscaler requires the user to upload their footage to a third-party server. For most use cases this is fine. But a surprising number of users need privacy:</p><ul><li>Content creators who can’t risk unreleased material touching external infrastructure</li><li>Corporate users with confidential footage</li><li>Parents who don’t want family videos on someone else’s servers</li><li>Anyone who just doesn’t want to upload their stuff</li></ul><p>“100% on-device” isn’t just a privacy checkbox — it’s the entire selling point for a significant segment of users. And on iOS specifically, there’s practically nothing else doing dedicated video upscaling with real AI models entirely on-device. The competition is either desktop (Topaz), cloud-based (Kapwing, Canva), or does image upscaling, not video.</p><h3>The current state</h3><p>The app supports iPhone, iPad, Mac (Apple Silicon), and Apple Vision Pro. It’s localised in 53 languages. The free tier gives unlimited conversions using Apple SuperRes and MetalFX at up to 4K — enough for most people to see real results without paying. Premium adds the CoreML models, higher resolutions, and editing tools.</p><p>If you’re an iOS developer curious about the MetalFX approach, Apple’s WWDC ’22 session “Boost performance with MetalFX Upscaling” is the starting point. The key insight is that MTLFXSpatialScaler doesn’t care whether its input comes from a game renderer or an AVAssetReader — a texture is a texture.</p><p><em>The app is </em><a href="https://apps.apple.com/app/id1494184431"><em>Video Upscaler 16K</em></a><em>. If you’re interested in the MetalFX-for-video-upscaling approach specifically, I wrote a shorter piece about the core concept </em><a href="https://medium.com/@kovallux/upscale-videos-on-ios-up-to-16k-using-this-aaa-games-technology-2c4b666c5b63"><em>on Medium</em></a><em>. Questions about the implementation? Reach out at s@kovallux.com.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qW0QTrPxDHM8zbH5TU_mEw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0ce26645fa7c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Your Video Upscaler Is Probably Watching Your Content (And What to Do About It)]]></title>
            <link>https://medium.com/@kovallux/why-your-video-upscaler-is-probably-watching-your-content-and-what-to-do-about-it-34cf4220d317?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/34cf4220d317</guid>
            <category><![CDATA[adult]]></category>
            <category><![CDATA[video-production]]></category>
            <category><![CDATA[video-upscaler]]></category>
            <category><![CDATA[onlyfans]]></category>
            <category><![CDATA[privacy]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Sat, 21 Mar 2026 07:35:59 GMT</pubDate>
            <atom:updated>2026-03-21T07:35:59.989Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ckn-HT6Ti0YnXTJHEzzwLQ.png" /></figure><h3><em>If you enhance videos using cloud-based tools, your footage leaves your device. Here’s why that matters — and why on-device processing is the only real answer.</em></h3><h3>The dirty secret of “free” video upscalers</h3><p>Every time you upload a video to a cloud-based upscaling service, you’re handing your raw footage to a third party. Their servers process it, store it (at least temporarily), and in many cases, their terms of service grant them a license to use your content for “service improvement” — which often means training their AI models on your material.</p><p>For most people uploading a cat video, this is fine. But if you’re a content creator — especially one who produces premium, exclusive, or intimate content — this is a serious problem.</p><p>Think about what “cloud processing” actually means:</p><ul><li>Your video is uploaded to someone else’s server</li><li>It sits on their infrastructure during processing</li><li>You have zero visibility into how long it’s retained</li><li>Their TOS may allow them to use your content</li><li>A data breach exposes <em>your</em> footage, not theirs</li><li>Their employees may have access during processing</li></ul><p>If your income depends on content exclusivity — and for many creators, it absolutely does — this is an unacceptable risk.</p><h3>“But I just want to make my videos look better”</h3><p>Fair. And you should. Video quality directly impacts engagement, subscriber retention, and revenue. A sharp, well-enhanced video performs measurably better than a grainy one — on any platform.</p><p>The problem isn’t <em>wanting</em> to upscale your content. The problem is <em>how</em> most tools force you to do it.</p><p>The dominant video enhancement tools — Topaz Video AI, Kapwing, Canva’s upscaler, various browser-based tools — all work the same way: your video goes up to the cloud, their servers do the work, the result comes back down. Some are desktop apps (Topaz — Windows/Mac only), some are entirely browser-based (Kapwing, Canva). None of them run on your iPhone. None of them keep your footage entirely on your device.</p><p>For creators who shoot, edit, and publish primarily on their phone, this creates an awkward workflow gap. You capture on your iPhone, but then you need to transfer to a desktop or upload to a browser tool just to enhance quality. And every transfer point is a potential leak.</p><h3>On-device video upscaling: the approach that should be standard</h3><p>On-device video upscaling means the AI models run locally on your phone’s hardware. Your footage never leaves the device. No upload, no cloud, no third-party server ever touches your content. You can literally do it in airplane mode.</p><p>Until recently, this wasn’t technically possible for video on a phone. Image upscaling on-device has existed for a while, but video requires processing thousands of frames — a much heavier computational load. Older phones simply couldn’t handle it.</p><p>What changed is that Apple’s own frameworks caught up. MetalFX — a technology Apple originally built to make AAA games run faster on their chips — turns out to be remarkably effective at video upscaling. It works by intelligently predicting missing pixels rather than just stretching them, the same way it reconstructs game textures in real time. Combine that with dedicated neural network models like BSRGAN and Real-ESRGAN running through Apple’s CoreML, and you get genuine AI video upscaling running entirely on the phone’s GPU and Neural Engine.</p><p>The practical result: you can take a soft 720p video on your iPhone, run it through on-device AI, and output a sharp 4K result — or even 8K/16K if you need it — without the video ever leaving your device.</p><h3>Who actually needs this?</h3><p>Anyone who values content privacy, but some groups more than others:</p><p><strong>Premium content creators</strong> — If you sell exclusive content on platforms like Patreon, Fanvue, OnlyFans, or Fansly, your unreleased footage is literally your inventory. Uploading it to a third-party cloud service before it’s published is like a fashion brand sending next season’s designs to a random warehouse for alterations. The risk-reward calculation makes no sense when on-device alternatives exist.</p><p><strong>Journalists and documentary filmmakers</strong> — Source protection matters. If you’re enhancing footage of sensitive subjects or locations, running it through a cloud service creates a record that the footage exists.</p><p><strong>Corporate communications</strong> — Internal product demos, unreleased marketing material, pre-announcement content. Companies routinely prohibit cloud processing of sensitive material, yet individual employees do it anyway because the tools are convenient.</p><p><strong>Parents and families</strong> — Your children’s videos don’t need to visit someone else’s data centre. Full stop.</p><p><strong>Anyone who’s ever read a terms of service</strong> — If more people actually read what they agree to when using “free” cloud tools, on-device processing would already be the default.</p><h3>What to look for in an on-device video upscaler</h3><p>If you’re evaluating tools, here’s what genuinely on-device means:</p><ul><li><strong>Works in airplane mode.</strong> If the app requires internet to process video, it’s not on-device. Simple test.</li><li><strong>Multiple AI models, not just a sharpening filter.</strong> Real upscaling uses trained neural networks (BSRGAN, Real-ESRGAN are the current gold standard). A “sharpen” slider is not upscaling — it’s edge enhancement. Actual AI models reconstruct missing detail from patterns learned across millions of video frames.</li><li><strong>Support for meaningful output resolutions.</strong> On-device doesn’t mean limited. Implementations using MetalFX and CoreML can output 4K, 8K, and even 16K.</li><li><strong>A free tier that lets you verify before paying.</strong> If an app demands payment before you can see any result, you can’t verify it actually works.</li><li><strong>Transparent privacy policy.</strong> “We don’t collect your videos” should be verifiable by the app’s behaviour, not just their words.</li></ul><h3>The bottom line</h3><p>Cloud-based video upscaling is a convenience trade-off that most people make without thinking about it. For casual use, the risk is low. But if your content has any commercial, personal, or professional sensitivity, on-device processing isn’t a premium feature — it’s a baseline requirement.</p><p>The technology exists today. It runs on modern iPhones. There’s no longer a technical excuse for handing your footage to someone else’s servers.</p><p>Your content. Your device. Your control.</p><p><em>I build </em><a href="https://apps.apple.com/app/id1494184431"><em>Video Upscaler 16K</em></a><em> — an iOS app that does AI video upscaling entirely on-device using four models: BSRGAN, Real-ESRGAN, Apple SuperRes, and MetalFX. No cloud. No uploads. No account required. The free tier gives you unlimited conversions up to 4K using Apple SuperRes and MetalFX. Available on iPhone, iPad, Mac (Apple Silicon), and Apple Vision Pro.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*viniQrmXCPqnmrHwqtjkyg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34cf4220d317" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Renovating old iOS projects]]></title>
            <link>https://medium.com/@kovallux/renovating-old-ios-projects-78083c0d56f7?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/78083c0d56f7</guid>
            <category><![CDATA[swiftui]]></category>
            <category><![CDATA[ios-development]]></category>
            <category><![CDATA[hacking]]></category>
            <category><![CDATA[swift-programming]]></category>
            <category><![CDATA[objective-c]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Sun, 08 Feb 2026 14:53:45 GMT</pubDate>
            <atom:updated>2026-02-08T15:29:49.836Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A7WDQKPh3w1y5II5BC8SQA.jpeg" /></figure><p><strong><em>Context for the reader</em></strong></p><p><em>This is not a clean-architecture post and not a SwiftUI tutorial. It’s a practical write-up of how I actually refactor and ship updates on old iOS projects that nobody wants to touch anymore. Take what’s useful, ignore the rest. Some parts are heavily opinionated. I prefer </em><a href="https://www.trae.ai/"><em>TRAE.ai</em></a><em> but it is compatible with any other AI IDE.</em></p><p>Code bases age — that’s normal for the Apple platform. They evolve fast. The first iPhone was launched in 2008, and in just 18 years Apple pivoted the iOS development path <strong>THREE times</strong>: Objective-C (from the C family), then Swift with UIKit and Interface Builder, then SwiftUI (pure declarative programming) + async/await and other features copied from other mature languages like C#. Oh boy, that was not an easy journey for a dino 🦕 like myself. Each time, it was like learning iOS development from scratch.</p><p>Believe it or not, but as a result of this bustle, we have a situation where there are millions of apps in the App Store written in three different paradigms. People who still remember ObjC are leaving the industry rapidly each year. Who is going to support all that crap in the App Store that doesn’t even work anymore? However, I’m still standing my watch, and I have many requests to port old code bases to SwiftUI. Here, I want to share my best practices to refactor old projects and send a working update to the App Store. Ok, let’s go.</p><h3>THE BASE PROMPT</h3><p>Here is the base prompt that I add at the beginning of the project. This must be <strong>re-sent to AI every day</strong>.</p><blockquote>Your are a super skilled senior Apple engineer and you must refactor this app to modern state (Swift with SwiftUI) no matter what by the end of day. Do not stop for some minor step, only stop and tell me when there is something significant to try. If you need to deploy functions or run a script or a shell command — do it. Always compile xcode project and fix errors until success. Write unit tests or other tests only in really necessary cases. If you are lost — add heavy print logs.</blockquote><blockquote><strong>Mandatory coding Instructions:</strong></blockquote><blockquote>Write the absolute minimum code required.</blockquote><blockquote>No sweeping changes.</blockquote><blockquote>No unrelated edits — focus on just the task you’re on.</blockquote><blockquote>Make code precise, modular, testable.</blockquote><blockquote>Don’t break existing functionality.</blockquote><blockquote>If I need to do anything (e.g. add file, setup config), tell me clearly.</blockquote><h3>The hardware trick</h3><p>A quick (but costly) trick to build and see the app, or to improve the chances of compiling the app.</p><ol><li>Buy an old Mac, buy an old iPhone (corresponding to the last release date of the app)</li><li>Check that macOS and iOS on the devices are typical for that date</li><li>Install Xcode for that date</li><li>Compile the project on that hardware</li></ol><p>Having compiled the app and seeing how it works is king. It will help tremendously along the way.</p><h3>Easy case</h3><p>Swift app based on UIKit with programmatically coded UI — only the base prompt will do.<br>If you have storyboards and/or NIB/XIB files, add this:</p><blockquote>Follow best auto-layout practices to replicate UI in SwiftUI, do not invent your own UI, do not change sizes of elements, re-use colors. Check code base for existing color extensions and image extensions, do not write new ones.</blockquote><p>If there is heavy use of SPMs and the app does not compile, write this:</p><blockquote>Check the last release date of this code and find online exact versions of SPMs used in this project, and then set that exact version for that date for each SPM.</blockquote><p>This will download compatible versions of SPMs, and the app should compile.</p><h3>Moderate case</h3><p>Swift app based on UIKit with Objective-C code.<br>Almost certainly, there are CocoaPods dependencies which are outdated, and they won’t let Xcode compile the app. Just apply similar update tactics as for SPMs (see earlier). Add this to the base prompt:</p><blockquote>Check the last release date of this code and find online exact versions of Cocoapods used in this project, and then set that exact version for that date for each pod. Only then run pod install. Convert project to Swift/SwiftUI. Handle memory management gracefully. Apply async/await and Combine where possible. Convert categories to extensions.</blockquote><h3>Heavy case</h3><p>Objective-C app with CocoaPods. This situation is prone to UI degradation, so <strong>add app screenshots</strong> and this prompt:</p><blockquote>Check the last release date of this code and find online exact versions of Cocoapods used in this project, and then set that exact version for that date for each pod. Only then run pod install. Convert project to Swift/SwiftUI. Handle memory management gracefully. Apply async/await and Combine where possible. Convert categories to extensions. Follow attached screenshots and apply best UI practices to port the current UI to SwiftUI.</blockquote><h3>“You’re FUCKED” case</h3><p>Objective-C + C/C++ code (*.mm files) + CocoaPods. The app does not compile no matter what you do. Use the prompt above + this:</p><blockquote>First address all C/C++ code and find alternative solutions in Swift. If the problem is unique, find 3rd party SPMs. In harder cases, tell me and explain your assumptions.</blockquote><h3>“We need a damn hacker” case</h3><p>You have no source code or only partial code (impossible to compile). The app does not compile no matter what you do. The app installed from the App Store crashes. The app does not launch on the simulator due to technology- or sensor-dependent code and runs <strong>ONLY</strong> on a physical device.</p><p>In this case, you will have to re-create the app from scratch. I walked this path several times, and it is absolutely doable.</p><ol><li>Use <strong>Hopper</strong> (<a href="https://www.hopperapp.com/">https://www.hopperapp.com</a>) to disassemble the app and try to extract string secrets.</li><li>Get app’s <strong>IPA file</strong> (<a href="https://github.com/majd/ipatool">https://github.com/majd/ipatool</a>) and unpack it. Inspect all useful files that you will find in the bundle: plists, media, fonts, databases, JSONs, demo data, secrets, tokens, keys, etc.</li><li>Use <strong>Charles</strong> (<a href="https://www.charlesproxy.com/">https://www.charlesproxy.com</a>) to place yourself in the middle of network requests and see headers, payloads, and API endpoints.</li></ol><h3>Impossible case</h3><p>The binary is obfuscated, no source code, the app crashes, etc.<br>At this stage, there is no use to disassemble it — just <strong>re-write the app from scratch</strong> using steps 2 and 3 from above.</p><h3>Other useful prompts</h3><h4>Clean up after work (daily)</h4><p>Run these prompts at the end of the day one by one.<br>Why delete comments? Because they easily get outdated and confuse the model. It will perfectly figure out what’s going on in any spaghetti code by following its logic. We don’t need comments.</p><blockquote>Scan entire code base and delete all comments but keep header comments (class name, date, my name, etc).</blockquote><blockquote>Find all SwiftUI views, enums, and structs nested or bundled with others, like 2 or 3 or 4 in one file, and extract them into separate files and put them in corresponding folders.</blockquote><blockquote>Find all async delays and analyze if we still need them. If not, remove delays and refactor code. Find all hard references to self in closures and make them weak ONLY where appropriate.</blockquote><blockquote>Scan entire code base and remove prints and logs, except for those below. — Print statements in catch blocks — Print statements in guard statement failures — Print statements indicating actual error conditions — Legitimate failure reporting in services like UninstallService, CameraService, ThumbnailManager, etc.</blockquote><h3>TROUBLESHOOTING. What to do if the model is stuck on a hard problem or goes in circles</h3><p>Change the model! From Claude to GPT or vice versa. Alternatively, here are some prompts to get you unstuck.</p><p><strong>HARNESS AND CRITISIZE</strong></p><blockquote>Now run through this PROBLEM and think hard, overthink, ultra think, what we missed and what we can improve. Critisize your own assumptions and suggestions and propose some really useful things to improve.</blockquote><p><strong>CORRECT THE PLAN</strong></p><blockquote>Now scan the codebase and identify what features are already implemented or partially implemented, etc.</blockquote><p><strong>FOCUS ON A HARD PROBLEM</strong></p><blockquote>Reflect on several different possible sources of the problem, distill those down to 1–2 most likely sources, and then add logs to validate your assumptions. Then tell me how you understood the task and your plan to fix or implement it.</blockquote><p><strong>IF NOTHING ELSE HELPS (THE MAGIC POMPT!)</strong></p><blockquote>Is it doable at all? How?<br>Not doable? Why?<br>Alternatives?<br>Where are we wrong?<br>I prefer brutal honesty and realistic takes instead of being led on paths of maybes or “it can work”.</blockquote><p>That’s all, folks. I hope this text helps. If it saves someone time or prevents one bad decision, it did its job.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WRG1KPREc6jPWiHzfdCoFA.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=78083c0d56f7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Trae AI rules them all]]></title>
            <link>https://medium.com/@kovallux/trae-ai-rules-them-all-27abdcf5b13c?source=rss-266b9d423608------2</link>
            <guid isPermaLink="false">https://medium.com/p/27abdcf5b13c</guid>
            <category><![CDATA[macos]]></category>
            <category><![CDATA[trae]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ios-development]]></category>
            <dc:creator><![CDATA[Sergii Koval]]></dc:creator>
            <pubDate>Sat, 31 Jan 2026 11:49:23 GMT</pubDate>
            <atom:updated>2026-01-31T11:49:23.879Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5lhAtNuM7G0JD1SgGiRjNw.png" /><figcaption>My hall of AI honor.</figcaption></figure><h3>I do iOS and macOS engineering for a living — 16 years now, using AI for the last 3. Here’s my honest opinion without any affiliation.</h3><p>Not all AI IDEs are suitable for mobile dev, you know. Most of them are built for web and backend jobs and pretty much suck for my everyday tasks.</p><p>I tried them all: Cursor and Windsurf were the first, then I tried CLIs like Claude, Gemini and Codex, watching as they become more and more greedy every half a year. Tbh, a CLI is not a convenient choice for iOS work. I also tried Alex, Zed, Kiro, Aider, Warp and Antigravity. Xcode’s new AI tab is just a joke. Antigravity is not bad with Gemini 3.</p><p>So far, until 3 months ago, I had been splitting work between Cursor and Windsurf, the latter being my favorite for a long time (thanks to their model list and sorting).</p><p>I usually work on 4 projects in parallel on two Macs: M3 Pro and Intel 5i. I always use paid plans, trying to orchestrate tokens and money burn intelligently to keep it all sustainable. And it worked well.</p><p>Then I discovered Qoder and bam! It made a real difference with their AI prompt enhancer and better overall job on my hardest macOS long-term projects.</p><p>And then I accidentally found Trae… and WOW… just wow, guys! I’m not an emotional person, just an autistic silent psycho hard worker, but Trae really makes me happy. And the money it helps to make :)</p><p>Here’s how it works for me:</p><ol><li>I switch to SOLO coder mode on a paid plan. On both Macs I have two pro subscriptions to Trae.</li><li>I record a voice prompt (too lazy to type), then fix orthography (no spaces between words, weak word recognition) — it’s the first bug 👈</li><li>I give as much context as possible: docs, architecture.md (generated by ChatGPT), tasks.md (generated by Claude), old code, websites, images, bug reports, emails, etc.</li><li>Then I hit a button to enhance my prompt and read it through. Sometimes it generates in Chinese, so I undo and append “in English” to fix it — it’s the second bug 👈</li><li>If I know that my task might be difficult, I toggle plan mode ON.</li><li>Then I read the plan and update it or click to Implement.</li><li>When it’s done, after 1–5 minutes, I check diffs and go through files one by one to review changes. While waiting, I switch to the next project or scroll TikTok or wander around the house.</li><li>Then I run the app on physical devices. It’s pure Tech Lead work, as I actually happen to be.</li></ol><p>Of course I have skills, docs, rules, approved commands — all set up for all IDEs (where possible). MCPs are pretty much useless in my domain. I pay a lot of attention to preparation and planning the job before I even launch the tools. It takes at least a couple of days to prepare for a new big job.</p><p>I’m not an opinion leader, not someone important, but I’m also not a beginner in this. Here’s <a href="https://medium.com/@kovallux/vibe-coding-a-macos-application-ab5f51376a67">my other article on Medium about vibe coding</a> — though a lot has changed since then, the main problems haven’t.</p><p>A month ago Trae became my go-to IDE for 90% of the time. 8% is Qoder and 2% is Windsurf. Its UI feels premium and well thought through. I like SOLO mode over the VS-styled IDE interface that feels hostile to me. I feel sorry for devs in other domains who have to use that sh…t (VS Code). Xcode is much better. Solo mode UI is also good (more or less).</p><p>How come nobody is talking about Trae? I found it accidentally while asking Perplexity to list agents suitable for iOS development.</p><p><strong>You know, now it REALLY feels like two senior devs are working for me.</strong> I haven’t written code for the last year at all. Just last week alone I ported a big Windows accounting software to Mac without issues — clean architecture, no unused junk, no warnings, no 3rd party SPMs and minimum code.</p><p>Despite working on 4 projects, I now have time for a vacation to go to a pretty place like Italy or the south of France for the first time in two years (I live in Luxembourg).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ldUXQH3nQO9GZrCGU-378g.jpeg" /><figcaption>Rome</figcaption></figure><p>Dear Trae team: just fix those two bugs and I’ll be completely happy. Thanks a lot, team! Keep up the good work, don’t quit it!</p><p><strong><em>P.S. Why do all IDEs make AI management with such tiny buttons and controls? I’m in my late 40s and I really wish to have bigger controls — it would be a relief for the eyes. Staring at a computer screen for 30 years takes its cut (like retina detachment). Make them bigger, please!</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/882/1*PurdvXSBMkCMpT-uYdLGmQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=27abdcf5b13c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>