<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Sherlabs Blog</title>
    <link>https://www.sherlabs.com/blog</link>
    <description>Notes from an independent app studio: how Snipo, AutoShare and our other apps are designed, built and shipped.</description>
    <language>en-au</language>
    <lastBuildDate>Wed, 11 Nov 2026 00:00:00 GMT</lastBuildDate>
    <atom:link href="https://www.sherlabs.com/blog/rss.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Introducing Odhi: Health Insight That Stays on Your Phone</title>
      <link>https://www.sherlabs.com/blog/introducing-odhi</link>
      <guid isPermaLink="true">https://www.sherlabs.com/blog/introducing-odhi</guid>
      <pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate>
      <dc:creator>Nilesh Jayanandana</dc:creator>
      <description>Odhi reads the sleep, heart and activity data already on your phone, sets it against your own usual, and shows what stands out, with a capybara for company. Nothing leaves the phone. Observations only, not a diagnosis.</description>
      <enclosure url="https://www.sherlabs.com/blog/introducing-odhi/cover.jpg" type="image/jpeg" length="0" />
      <content:encoded><![CDATA[<img src="https://www.sherlabs.com/blog/introducing-odhi/cover.jpg" alt="Odhi, a capybara wearing a flower garland, with Pip the bird" /><figure><img src="https://www.sherlabs.com/blog/introducing-odhi/cover.jpg" alt="Odhi, a capybara with a flower tucked by his ear, with Pip the bird" width="1400" height="787" loading="lazy" decoding="async" /><figcaption>Odhi, a capybara with a flower tucked by his ear, with Pip the bird</figcaption></figure>
<p>Most of us now carry a quiet record of our own health. Sleep from a watch. Heart rate. Steps. A headache we logged once and forgot. It sits in a health app, in lines and rings, and mostly it doesn&#39;t tell you anything.</p>
<p><strong>Odhi</strong> is my attempt at making that record useful without making it anyone else&#39;s business. It reads the data already on your phone, sets it against <em>your own</em> usual, and shows what stands out. Nothing leaves the phone.</p>
<p>It isn&#39;t on the App Store yet. This is what it is, how it behaves, and what it deliberately doesn&#39;t do.</p>
<h2>Against your usual, not against everyone</h2>
<p>A sleep score that compares you with a population average isn&#39;t very helpful. You are not the average. A seven-hour night might be ordinary for you, or a short one.</p>
<p>So Odhi only compares you with you. It looks at recent weeks against your earlier weeks: sleep shorter than usual, bedtimes drifting later, activity slowly falling, a resting heart rate that has stayed up for several days. Each note comes with its numbers, how much data was behind it, and what it can&#39;t tell you. Some days there&#39;s nothing to say, and Odhi says that too.</p>
<figure><img src="https://www.sherlabs.com/blog/introducing-odhi/today.webp" alt="Odhi's Today screen with sample data: a capybara resting in a pond above a card that reads &quot;A pattern is worth a look&quot;" width="600" height="1304" class="post-shot" loading="lazy" decoding="async" /><figcaption>Odhi's Today screen with sample data: a capybara resting in a pond above a card that reads &quot;A pattern is worth a look&quot;</figcaption></figure>
<p>The screen above uses made-up sample data. That&#39;s what demo mode shows, and it&#39;s labelled as such.</p>
<h2>Observations only</h2>
<p>Odhi is not a medical app. It gives observations, not a diagnosis, and it won&#39;t tell you that you are well or unwell. When it says something is worth mentioning to your GP, that&#39;s a prompt to ask, not a conclusion.</p>
<p>I wrote down a few rules before building, and they still hold:</p>
<ul>
<li>No health score out of 100. No streaks. No guilt about the night you didn&#39;t track.</li>
<li>No red alarm colours, and no words like &quot;risk&quot; or &quot;abnormal&quot;.</li>
<li>Wording like &quot;associated with&quot; and &quot;may&quot;, never &quot;because&quot;.</li>
<li>No advice about medicine, ever.</li>
</ul>
<h2>A capybara called Odhi, and a bird called Pip</h2>
<p>Health data can feel heavy, so Odhi comes with company. A small capybara and a bird called Pip drift around the Today screen. They never look worried, never scold, and never mention a condition. If they aren&#39;t your thing, one switch in Settings makes them disappear, and everything else stays the same. There&#39;s also a home-screen widget that shows a picture and a single short sentence, with no numbers on it.</p>
<h2>Private, properly</h2>
<p>Odhi has no account and no server. Your records live in an encrypted database on your phone. The release build can&#39;t use the internet at all, and the app has no analytics, no ads and no crash reporting. Odhi reads Apple Health and writes to it only if you turn on the optional setting to save lab results there.</p>
<p>You can export everything, erase everything, or delete the app and take it all with you. The full details are in the <a href="https://www.sherlabs.com/privacy/odhi">privacy policy</a>.</p>
<h2>The experimental parts</h2>
<p>A few things I&#39;m trying are not ready to be called finished. They are off by default and kept out of the App Store version:</p>
<ul>
<li>an optional overnight listen that measures snoring, where sound is turned into a few numbers and never stored or sent,</li>
<li>a migraine heads-up that only appears if a test on your own history beats simple guesses,</li>
<li>a gentle &quot;pace&quot; word for whether your routines are steadier, about the same or less steady.</li>
</ul>
<p>They have been tried on made-up data and in internal test builds, and no further. They are not checked against sleep studies or clinical measurements. I&#39;ve written up what they draw on, what they don&#39;t claim, and how they work, on the <a href="https://www.sherlabs.com/odhi/research">research pages</a>, and I explain the thinking in <a href="https://www.sherlabs.com/blog/how-odhi-reads-research">How Odhi reads research</a>.</p>
<h2>What&#39;s next</h2>
<p>I&#39;d like to get Odhi onto the App Store, and I&#39;d like to hear from you first if you&#39;d like to try it. You can <a href="https://www.sherlabs.com/contact?about=Odhi">tell me through the contact form</a> and I&#39;ll get back to you. Please don&#39;t send health information.</p>
]]></content:encoded>
    </item>
    <item>
      <title>How Odhi Reads Research: What We Cite, and What We Don’t Claim</title>
      <link>https://www.sherlabs.com/blog/how-odhi-reads-research</link>
      <guid isPermaLink="true">https://www.sherlabs.com/blog/how-odhi-reads-research</guid>
      <pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate>
      <dc:creator>Nilesh Jayanandana</dc:creator>
      <description>Only three papers made it into the pile we have read in full. How Odhi sorts research into read, summary only and left out, and why several features cite nothing at all.</description>
      <enclosure url="https://www.sherlabs.com/blog/how-odhi-reads-research/cover.jpg" type="image/jpeg" length="0" />
      <content:encoded><![CDATA[<img src="https://www.sherlabs.com/blog/how-odhi-reads-research/cover.jpg" alt="Odhi, a capybara, standing up to look at something with Pip the bird nearby" /><figure><img src="https://www.sherlabs.com/blog/how-odhi-reads-research/cover.jpg" alt="Odhi, a capybara, standing up to look at something with Pip the bird nearby" width="1400" height="787" loading="lazy" decoding="async" /><figcaption>Odhi, a capybara, standing up to look at something with Pip the bird nearby</figcaption></figure>
<p>Health apps love a citation. A small superscript number, a journal name, and suddenly a rough guess feels settled. I didn&#39;t want Odhi to do that. So when I wrote its research pages, I set myself a stricter rule than I expected to need.</p>
<h2>Three piles</h2>
<p>Everything I found went into one of three piles.</p>
<p><strong>Read in full.</strong> I read the whole paper. Only these are cited as support.</p>
<p><strong>See also, summary only.</strong> I know the paper from its abstract or a summary. I list it so you can go and read it yourself, but I quote no figures from it, because I can&#39;t vouch for them.</p>
<p><strong>Left out.</strong> Anything I couldn&#39;t check at all, or whose reference I couldn&#39;t confirm. These don&#39;t appear.</p>
<p>When I sorted what I had, the result was humbling. Only three papers made the first pile, and for two of Odhi&#39;s features, none did.</p>
<h2>What that looks like in practice</h2>
<p>For the optional snoring feature, the two papers I&#39;ve read in full are about wearables worn by people being assessed for sleep apnoea, not about a phone listening to a room. They report how well a wrist or finger device agreed with a sleep test. They&#39;re useful for a small supporting signal and for a reminder of how much results depend on who was studied. They say nothing about whether a phone on a bedside table can do the same job, so the page says that plainly.</p>
<p>For the migraine heads-up, I have read none of the related papers in full. The page says &quot;Cited as support: none&quot;, lists the papers I know as summary only, and notes that results in this area look modest and uneven. That&#39;s why the feature shows itself only when a test on your own history beats two naive guesses, and why many people will never see it.</p>
<p>For the fitness estimate, I read the paper behind a published model. Then I found I couldn&#39;t verify the equations from a secondary summary, so Odhi holds none of them. The feature is blocked until someone transcribes the equations from the paper and checks them.</p>
<p>For the everyday insights, no clinical paper is relied on at all. They use ordinary statistics on your own data.</p>
<h2>Marking every threshold</h2>
<p>A second habit: every number in the <a href="https://www.sherlabs.com/odhi/open-method">open method</a> is labelled as cited, summary only, a design value, or a heuristic. A heuristic is a rule of thumb I chose, with no study behind it. There are many of those, and I&#39;d rather you see them than assume they were borrowed from somewhere official.</p>
<h2>Not claiming too much</h2>
<p>Odhi&#39;s experimental features carry one status line everywhere they appear: observations only, not a diagnosis, not clinically validated against sleep studies, experimental. I mean each part of it. They have been tried on made-up data and in internal tests. They haven&#39;t been run against real sleep-study or migraine data.</p>
<p>I&#39;m one person, not a research group. If you work in a university, hospital or sleep lab and want to test, challenge or improve any of this, the <a href="https://www.sherlabs.com/odhi/open-method">open method page</a> has an outline for testing it properly and a way to reach me. I&#39;d rather be shown a flaw than keep one.</p>
]]></content:encoded>
    </item>
    <item>
      <title>Building OpenHarness: A Local-First Home for a Fleet of Coding Agents</title>
      <link>https://www.sherlabs.com/blog/building-openharness</link>
      <guid isPermaLink="true">https://www.sherlabs.com/blog/building-openharness</guid>
      <pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate>
      <dc:creator>Nilesh Jayanandana</dc:creator>
      <description>OpenHarness runs coding agents side by side on your own machine, each task in its own git worktree with a chat, a terminal and a diff. Why it is local-first, how it was built, and what is still unfinished.</description>
      <enclosure url="https://www.sherlabs.com/blog/building-openharness/cover.jpg" type="image/jpeg" length="0" />
      <content:encoded><![CDATA[<img src="https://www.sherlabs.com/blog/building-openharness/cover.jpg" alt="OpenHarness: run coding agents side by side, with the app window under the Sherlabs wordmark" /><figure><img src="https://www.sherlabs.com/blog/building-openharness/cover.jpg" alt="OpenHarness: run coding agents side by side" width="1400" height="787" loading="lazy" decoding="async" /><figcaption>OpenHarness: run coding agents side by side</figcaption></figure>
<p>Once you trust a coding agent with a real task, you want to give it three. Then the problems start. They share one working folder and trip over each other&#39;s changes. Each lives in its own terminal tab. One of them has been waiting on a permission prompt for twenty minutes and you did not notice. And a few agents running builds and tests at once can make a laptop crawl.</p>
<p><strong>OpenHarness</strong> is my answer: a desktop app for running coding agents side by side. Add a git repository, start a workspace for each task, and each one gets its own git worktree, its own chat, a terminal and a diff. Everything runs on your computer. It is an early preview, and this post is as much about what is not finished as what is.</p>
<h2>What works today</h2>
<ul>
<li><strong>Projects and workspaces.</strong> Every workspace is a git worktree on its own branch, so agents on the same repository stay out of each other&#39;s files.</li>
<li><strong>Claude Code chat with approvals.</strong> A structured chat instead of a raw terminal. Tool calls show up as cards, and anything that needs permission waits for you.</li>
<li><strong>A terminal in every workspace.</strong></li>
<li><strong>A diff view</strong> of what the agent changed against HEAD.</li>
</ul>
<figure><img src="https://www.sherlabs.com/blog/building-openharness/workspace.webp" alt="OpenHarness with sample projects. Illustrative data." width="1600" height="1000" loading="lazy" decoding="async" /><figcaption>OpenHarness with sample projects. Illustrative data.</figcaption></figure>
<p>That screenshot uses made-up projects, and it also shows a Fleet entry that is not finished. I would rather say so than let it imply more.</p>
<h2>Why local-first</h2>
<p>I wanted a tool I could trust with my own code without asking anything of anyone else. So OpenHarness has no account and no telemetry. Its data lives in a <code>.openharness</code> folder in your home folder. The agents are ordinary command-line tools that you install and sign in to yourself, and OpenHarness supervises them as processes. It never asks for your provider login.</p>
<p>That decision had a consequence. OpenHarness does not download agent binaries from servers of its own. It looks for the CLIs you already have and tells you which it found. The one thing it asks my website for is whether a newer version exists, and you can turn that check off.</p>
<h2>Built with a fleet of agents</h2>
<p>Most of OpenHarness was written by coding agents working in parallel, each in its own worktree, with me reviewing. That is exactly the workflow the app exists for, so I felt the missing pieces as I went: the approval I missed, the machine slowing down, the worktrees piling up.</p>
<p>It is a Tauri app. The shell and the process supervision are Rust, the interface is React and TypeScript, a local SQLite database holds the state, and the terminal is a real pseudo-terminal.</p>
<h2>What is not finished</h2>
<p>These are built or half-built in the repository, and not in a release yet:</p>
<ul>
<li>A <strong>fleet dashboard</strong> with CPU, memory, network and disk use for each agent.</li>
<li>A <strong>governor</strong> that pauses or lowers the priority of agents so the machine never freezes.</li>
<li>A <strong>skills manager</strong>, and <strong>history</strong> to browse and resume past sessions.</li>
<li><strong>Worktree cleanup</strong> with safety backups.</li>
<li><strong>Optional memory</strong> through agentmemory, running on your own machine.</li>
<li>A <strong>skip-permissions mode</strong>, with warnings.</li>
</ul>
<p>Planned, but not started: adapters for <strong>Codex, OpenCode, Cursor and Gemini</strong>, and <strong>sandboxed runs</strong> through OpenSandbox.</p>
<h2>What it will not do</h2>
<ul>
<li><strong>It is an early preview.</strong> Expect rough edges, and expect things to change.</li>
<li><strong>It does not sandbox the agents.</strong> Today they run on your machine with your permissions. Read what an agent asks to do before you approve it.</li>
<li><strong>Only Claude Code works today.</strong> The other agents are plans, not features.</li>
<li><strong>Windows is experimental.</strong> The first builds are Mac only, and a Windows installer, when there is one, is not code-signed.</li>
<li><strong>It is unofficial.</strong> It is not affiliated with Anthropic.</li>
</ul>
<h2>Get it</h2>
<p>OpenHarness is free to <a href="https://www.sherlabs.com/openharness#download">download for Mac</a>. The source code is private for now. If something breaks, or you have an idea, please use the <a href="https://www.sherlabs.com/openharness#feedback">feedback form</a>.</p>
]]></content:encoded>
    </item>
    <item>
      <title>Figmake Sync: Edit Your Figma Make Project With Your Own AI</title>
      <link>https://www.sherlabs.com/blog/building-figmake-sync</link>
      <guid isPermaLink="true">https://www.sherlabs.com/blog/building-figmake-sync</guid>
      <pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate>
      <dc:creator>Nilesh Jayanandana</dc:creator>
      <description>Figma Make only lets you edit through Figma’s own AI, and its per-seat credits burn up fast. Figmake Sync pulls your project onto your computer so you can describe changes to the Claude or ChatGPT you already pay for, show it screenshots or clicked elements, review the changes, merge with Figma and send them back.</description>
      <enclosure url="https://www.sherlabs.com/blog/building-figmake-sync/cover.jpg" type="image/jpeg" length="0" />
      <content:encoded><![CDATA[<img src="https://www.sherlabs.com/blog/building-figmake-sync/cover.jpg" alt="Edit Figma Make with your own AI: Figma Make code, a local folder, Claude Code or Codex, and back to Figma, under the Sherlabs wordmark" /><figure><img src="https://www.sherlabs.com/blog/building-figmake-sync/cover.jpg" alt="Figmake Sync: edit Figma Make with your own AI" width="1400" height="787" loading="lazy" decoding="async" /><figcaption>Figmake Sync: edit Figma Make with your own AI</figcaption></figure>
<p>Figma Make builds a working prototype from a prompt. After that, every change goes through Figma&#39;s own AI, and that AI runs on Figma&#39;s AI credits. Credits come per seat and are capped every month: on Figma&#39;s pricing page today, a Professional full seat includes 3,000 AI credits a month, Organization 3,500 and Enterprise 4,250, and a Dev or Collab seat only 500. More means buying an add-on. We found those credits burn up fast. Small tweaks, a colour, a label, a total that was off, each spent credits, and the cost added up quickly.</p>
<p>Meanwhile we already pay for Claude and ChatGPT. They are good at exactly this kind of work, but they have no way into your project.</p>
<p><strong>Figmake Sync</strong> gives them one. It is a free desktop app for Mac and Windows. It brings your project onto your computer, lets you change it in a chat panel next to a live preview, and sends the result back to Figma. A sidebar holds all your projects.</p>
<h2>How it works</h2>
<ol>
<li><strong>Open your link.</strong> Paste your Figma Make link. The first time, you sign in to Figma and pick Claude or ChatGPT.</li>
<li><strong>Ask for a change.</strong> Type what you want in the chat, like &quot;make the button blue&quot;. The reply streams in, the preview updates, and each change shows up as a card with Undo and View changes.</li>
<li><strong>Send it to Figma.</strong> When you like it, press Send to Figma.</li>
</ol>
<h2>Why we built it</h2>
<p>We build our own prototypes in Figma Make. A small fix meant a prompt to Figma&#39;s AI, and a bigger one meant a long session of them. The credits went fast, and the cost of Figma&#39;s AI became one of our biggest reasons to look for another way. Our own AI, already paid for, sat idle next to a project it could not reach.</p>
<p>The code is visible in Figma, but you can only edit it by hand there. So the choice was spending Figma&#39;s credits or typing every change yourself. Figmake Sync is the third option: the same project, edited by the AI subscription you already have, with no Figma AI credits spent.</p>
<p>As far as we know, Figma Make has no public way to write changes back. So Figmake Sync does what you would do: it opens Figma in a window on your computer, using your own sign-in. We never see your password.</p>
<h2>What it does now</h2>
<p>A lot has changed since the first version. Most of it came from using it on real projects and watching it fail in ways I had not planned for.</p>
<h2>It brings its own Node</h2>
<p>The app now ships a sandboxed Node.js and pnpm and uses those to install and run your project. Nothing is installed on your computer, and an existing Node setup is left alone.</p>
<p>Previews are steadier too. Each project gets its own free port, and the app cleans up after itself, so you do not end up with preview processes still running after you close it.</p>
<h2>A preview that works like a browser</h2>
<p>The preview now has a browser bar with tabs, back, forward, reload and an address bar. It only opens pages on your own computer. It is there because checking a change often means clicking through a few pages, not looking at one.</p>
<h2>Watching the AI work</h2>
<p>While the AI works, the chat shows a step timeline: files it reads, edits and creates, checks it runs, any helper agents, and a Tasks card for longer jobs. You can stop it at any point. When it finishes, the steps fold into one line and a result card lists the files changed, with Undo and View changes.</p>
<figure><img src="https://www.sherlabs.com/blog/building-figmake-sync/chat.webp" alt="The chat with a step timeline, a screenshot and a picked element" width="1600" height="1040" loading="lazy" decoding="async" /><figcaption>The chat with a step timeline, a screenshot and a picked element</figcaption></figure>
<h2>Show it, do not describe it</h2>
<p>You can paste, drop or attach screenshots (up to five per message). There is also an element picker: press the target button in the preview bar, click part of the page, and it becomes a chip in the chat with the element&#39;s tag, text, styles and a cropped picture. The AI gets those details labelled as page content. I have not yet tested the picker on React newer than 18, on shadow DOM or inside iframes, so expect gaps there.</p>
<h2>Models, sign-in and the sidebar</h2>
<ul>
<li><strong>Model switching.</strong> A chip under the chat box opens a model menu for the AI you chose, and lets you switch between Claude and ChatGPT (Codex). It applies from your next message.</li>
<li><strong>Signing in.</strong> If Claude or ChatGPT is signed out, the app opens your browser and carries on once you are done.</li>
<li><strong>A sidebar that gets out of the way.</strong> The toggle moves into the left bar (Cmd or Ctrl+B) and collapses projects to a narrow rail. The preview defaults to Fit, which measures the page and scales an app that is too wide so it does not scroll sideways.</li>
<li><strong>Projects.</strong> Every project stays in the sidebar with its own chat. Switch between them freely.</li>
</ul>
<h2>Review before you send</h2>
<p>Send to Figma no longer sends blindly. It opens a list of every changed file with the lines added and removed. You can untick a file to leave it out, and open a diff (Monaco, the editor from VS Code) to read the exact change. Files you chose not to send are listed in Settings, where you can turn them back on.</p>
<figure><img src="https://www.sherlabs.com/blog/building-figmake-sync/review.webp" alt="The review screen with a diff open" width="1600" height="1040" loading="lazy" decoding="async" /><figcaption>The review screen with a diff open</figcaption></figure>
<h2>Sync with Figma</h2>
<p>If the project changed in Figma, press Sync with Figma. You have two options:</p>
<ul>
<li><strong>Merge</strong> brings Figma&#39;s changes into yours, line by line.</li>
<li><strong>Discard my changes</strong> throws yours away and uses Figma&#39;s version.</li>
</ul>
<p>When both sides edited the same lines of a file, the app cannot combine them. It asks you to pick per file, and shows the two versions side by side first. A backup is saved before either option runs.</p>
<figure><img src="https://www.sherlabs.com/blog/building-figmake-sync/merge.webp" alt="Choosing which version of a file to keep" width="1600" height="1040" loading="lazy" decoding="async" /><figcaption>Choosing which version of a file to keep</figcaption></figure>
<h2>A send that stops when it should</h2>
<p>Sending works by driving Figma&#39;s own screens, and those screens do not always cooperate. I spent a while on that.</p>
<p>The app now plans the whole send before it writes anything. It checks each file after it is sent, and does a final check at the end. If something does not match, it stops and says why in plain words, instead of pressing on and leaving half a project in Figma. A progress card shows the file being sent, how many are done and the percent.</p>
<figure><img src="https://www.sherlabs.com/blog/building-figmake-sync/sending.webp" alt="The progress card while sending" width="1600" height="1040" loading="lazy" decoding="async" /><figcaption>The progress card while sending</figcaption></figure>
<h2>Smaller things</h2>
<ul>
<li>Rename a project.</li>
<li>New projects go in <code>~/figmake-sync</code>. iCloud Documents made first reads and previews very slow, so Documents is no longer the default. You can move the folder in Settings.</li>
<li>When Figma renames a control, the app can ask your connected AI to find it again. This covers most steps, not all of them.</li>
<li>A Support Sherlabs button sits at the bottom of the sidebar. It opens the <a href="https://www.sherlabs.com/support-us">support page</a>. The app stays free either way.</li>
</ul>
<h2>Safe by default</h2>
<ul>
<li>Every change is backed up first, so each change can be undone.</li>
<li>If Figma has newer changes when you send, you choose <strong>Keep mine</strong> or <strong>Use Figma&#39;s</strong>.</li>
<li>The AI can only edit the files in your project folder.</li>
</ul>
<h2>What it will not do</h2>
<ul>
<li><strong>It is unofficial.</strong> It is not affiliated with Figma, and Figma has not reviewed it. If Figma changes its screens, a step can break. The app can try to repair a step using your connected AI, but not every step.</li>
<li><strong>It cannot see your Figma canvas.</strong> The AI works on your project&#39;s code, not on your frames.</li>
<li><strong>ChatGPT support is experimental.</strong> Claude is the one we use every day.</li>
<li><strong>Your project&#39;s code is not sandboxed.</strong> The app brings its own Node and pnpm, but your project&#39;s own code still runs on your computer, so use projects you trust.</li>
<li><strong>The send checks are newly tested.</strong> I tried them against a fake Figma explorer and a live read-only probe, not a full live end-to-end send on every kind of project. The Windows installer is unsigned and I have not run it on a real Windows machine.</li>
<li><strong>The element picker is new.</strong> Not yet tested on React newer than 18, shadow DOM or iframes.</li>
</ul>
<h2>If something goes wrong</h2>
<p>When a step fails, press <strong>Report</strong> next to Try again. It opens a GitHub issue in your browser. Nothing is sent until you review it and press Submit. It holds no code, prompts or project names.</p>
<h2>Get it</h2>
<p>Figmake Sync is free for <a href="https://www.sherlabs.com/products/figmake-sync#download">Mac and Windows</a>, and the source is on <a href="https://github.com/sherlabs/figmake-sync" target="_blank" rel="noopener noreferrer">GitHub</a>.</p>
]]></content:encoded>
    </item>
    <item>
      <title>Designing the Sherlabs Logo: One Stroke, Two Dots and a Single Blue</title>
      <link>https://www.sherlabs.com/blog/designing-the-sherlabs-logo</link>
      <guid isPermaLink="true">https://www.sherlabs.com/blog/designing-the-sherlabs-logo</guid>
      <pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate>
      <dc:creator>Nilesh Jayanandana</dc:creator>
      <description>The thinking behind the Sherlabs mark, with the real numbers: a lowercase wordmark on a 5.56 : 1 canvas, one 43-unit stroke, optical overshoot, an S that ends in a dot, an off-centre icon and one periwinkle blue.</description>
      <enclosure url="https://www.sherlabs.com/blog/designing-the-sherlabs-logo/cover.jpg" type="image/jpeg" length="0" />
      <content:encoded><![CDATA[<img src="https://www.sherlabs.com/blog/designing-the-sherlabs-logo/cover.jpg" alt="The Sherlabs wordmark on its type guides" /><figure><img src="https://www.sherlabs.com/blog/designing-the-sherlabs-logo/cover.jpg" alt="The Sherlabs wordmark on its type guides" width="1400" height="787" loading="lazy" decoding="async" /><figcaption>The Sherlabs wordmark on its type guides</figcaption></figure>
<p>A studio logo has an awkward job. It has to sit quietly in a 16-pixel browser tab, hold its own as a hero across a laptop screen, and look like it belongs next to apps as different as a reel organiser and a family photo tool. It can&#39;t lean on one product&#39;s look, because there will be more products.</p>
<p>The Sherlabs mark answers that with very little: a lowercase wordmark drawn with one stroke weight, two dots, and one colour. This post walks through the thinking behind each of those choices, with the real numbers from the master SVG.</p>
<h2>Lowercase, on purpose</h2>
<p>Every letter in the wordmark is lowercase, including the first. Capitals read as formal: a bank, a law firm, an institution. Sherlabs makes small apps for personal things like saved recipes and photos of your kids, so the name should read the way you&#39;d type it to a friend.</p>
<p>Lowercase also gives the mark a recognisable outline. The ascenders of the <strong>h</strong>, <strong>l</strong> and <strong>b</strong> rise out of the word at regular intervals and make a skyline you can spot even when the logo is too small to read.</p>
<figure><img src="https://www.sherlabs.com/blog/designing-the-sherlabs-logo/anatomy.svg" alt="Wordmark anatomy: ascender, x-height, baseline and overshoot" width="1400" height="600" loading="lazy" decoding="async" /><figcaption>Wordmark anatomy: ascender, x-height, baseline and overshoot</figcaption></figure>
<p>The proportions are set for that. The wordmark is drawn on a 1450 × 261 unit canvas, a 5.56 : 1 rectangle, wide enough to sit comfortably in a header and still be legible on a phone. The x-height is 184 units against an ascender of 257, a ratio of 0.72. That&#39;s a big x-height, the same trick friendly interface typefaces use: the bodies of the letters stay large, so the word stays readable when it&#39;s scaled down.</p>
<h2>One stroke weight</h2>
<p>Every vertical in the mark, from the stem of the <strong>h</strong> to the <strong>l</strong> to the back of the <strong>b</strong>, is exactly 43 units wide. The curves are drawn to the same weight. There are no thick-and-thin contrasts, no serifs, no flourishes.</p>
<p>A single weight is what lets the logo shrink. Contrast in a letterform is the first thing to disappear at small sizes: thin strokes break up into grey mush while thick ones survive. With one weight throughout, the whole word degrades evenly. At 43 units against a 184-unit x-height, the stroke is 23% of the letter height, heavy enough to feel confident without filling in the counters.</p>
<h2>Open letterforms</h2>
<p>Look closely and most letters are a little open. The stem of the <strong>b</strong> stops short of the baseline, leaving a small notch where it would meet the base of the bowl. The <strong>a</strong> is a single-storey form with a small spur kicking out at the bottom right. The <strong>r</strong> is just a stem and a quarter arc. The <strong>e</strong> ends on a clean horizontal cut instead of curling back on itself.</p>
<p>Open shapes keep the counters (the spaces inside letters) generous, which again helps at small sizes. They also give the mark a sense of motion, as if each letter flows into the next. It&#39;s a small, quiet way of saying the work is never quite finished: these are products we keep shipping.</p>
<h2>Optical corrections</h2>
<p>Geometry alone doesn&#39;t look right, so the mark cheats in the places your eye expects it to.</p>
<p>Round letters overshoot. The bowls of the <strong>e</strong> and <strong>a</strong> dip 4 units below the baseline, and the curved tops of the <strong>h</strong> and <strong>e</strong> rise 4 units above the flat top of the <strong>a</strong>. Without that, round shapes look shorter than flat ones next to them, because a curve touches a line at only one point.</p>
<p>The spacing isn&#39;t uniform either. The gaps between letters range from 25 units (between the <strong>a</strong> and <strong>b</strong>) to 33 (between the <strong>h</strong> and <strong>e</strong>). Each pair is spaced by eye, so that the <em>areas</em> of white between letters feel even, not the distances.</p>
<h2>The two dots</h2>
<p>The most distinctive part of the mark is what happens to the <strong>s</strong>.</p>
<figure><img src="https://www.sherlabs.com/blog/designing-the-sherlabs-logo/two-dots.svg" alt="The first S carries the blue dot, the last S an ink dot" width="1400" height="600" loading="lazy" decoding="async" /><figcaption>The first S carries the blue dot, the last S an ink dot</figcaption></figure>
<p>The S is drawn without its lower curl. Where the stroke would normally sweep back to the left at the bottom, it stops, and a dot sits in its place. The dot is 54 units across, about 1.26 times the stroke, so it reads as a deliberate shape rather than a full stop that drifted in.</p>
<p>Sherlabs starts and ends with an <strong>s</strong>, and the same S is used at both ends: the final one is the exact same path, moved 1,278 units to the right. What changes is the dot.</p>
<ul>
<li>The first dot is <strong>blue</strong>. It&#39;s the spark at the start of the name, the only colour in the whole mark.</li>
<li>The last dot is <strong>ink</strong>, the same colour as the letters. It closes the word like a full stop.</li>
</ul>
<p>Start with an idea, finish the thing. That&#39;s the closest the logo gets to a slogan.</p>
<h2>The icon</h2>
<p>The app icon and favicon keep only the opening S and its blue dot.</p>
<figure><img src="https://www.sherlabs.com/blog/designing-the-sherlabs-logo/icon-grid.svg" alt="The icon on its grid" width="1400" height="600" loading="lazy" decoding="async" /><figcaption>The icon on its grid</figcaption></figure>
<p>It sits on a 272 × 272 tile with a 48-unit corner radius, 17.6% of the side: clearly soft, a little tighter than an iPhone app icon, so it looks at home beside them without pretending to be one. The S is 172 units wide, 63% of the tile, which leaves enough margin for the rounded corners to breathe.</p>
<p>The S isn&#39;t vertically centred. It has 35 units above and 45 below. The dot and the heavier lower half of the letter weigh the shape down, so the mark is nudged upward to sit on the tile&#39;s <em>optical</em> centre. Mathematically centred, it looks like it&#39;s sinking.</p>
<p>The single stroke weight earns its keep here. At 43 units on a 272 tile, the stroke is 16% of the icon. In a 16-pixel browser tab that&#39;s still a 2.5-pixel line, so the S stays an S.</p>
<h2>A single blue</h2>
<figure><img src="https://www.sherlabs.com/blog/designing-the-sherlabs-logo/colours.svg" alt="The brand colours and their contrast" width="1400" height="560" loading="lazy" decoding="async" /><figcaption>The brand colours and their contrast</figcaption></figure>
<p>The mark uses one colour, and uses it once.</p>
<ul>
<li><strong>Logo blue, #728EFB.</strong> A light periwinkle: hue 228°, very saturated, fairly light. It&#39;s friendly rather than corporate, and it doesn&#39;t belong to any one product. The apps keep their own colours, like AutoShare&#39;s coral, and this blue sits comfortably beside them.</li>
<li><strong>Ink, #1A1A19.</strong> Not pure black. A very slightly warm near-black, which looks softer on screen than #000.</li>
<li><strong>Paper, #FAFAFA.</strong> An off-white for the page, so white cards and the light icon tile can still stand off it.</li>
</ul>
<p>The logo blue has one limitation, and it was a deliberate trade. It measures 3.01 : 1 against white. That passes the WCAG threshold for graphics, so the dot is fine on light backgrounds, but it&#39;s too light for small text. So the website doesn&#39;t use it for interface text. Text selection, focus rings and the small square markers on section labels use a deeper indigo, <strong>#6366F1</strong>, at 4.47 : 1 on white. On dark backgrounds the logo blue does much better, at 6.58 : 1 against near-black.</p>
<p>Keeping the logo blue for the logo, and a working blue for the interface, means the brand colour stays special. It only ever shows up as that one dot.</p>
<h2>Drawn, not typed</h2>
<p>The final mark contains no font. Every letter is an SVG path, and the whole wordmark is 1.7 KB in both its light and dark versions. Nothing depends on a typeface being installed or loading in time, and there&#39;s nothing to substitute if it doesn&#39;t.</p>
<p>There are four files, and that&#39;s the whole kit:</p>
<ul>
<li><code>sherlabs-wordmark-dark.svg</code> and <code>sherlabs-wordmark-light.svg</code>: ink or white letters on a transparent background, both with the blue dot.</li>
<li><code>sherlabs-icon-light.svg</code> and <code>sherlabs-icon-dark.svg</code>: the S and dot on a white or black tile.</li>
</ul>
<p>The rules for using them are short. Never stretch the 5.56 : 1 rectangle. Keep at least one dot&#39;s width of clear space on every side. Never recolour the first dot, and never add a second colour.</p>
<h2>What we&#39;d tell anyone designing a small brand</h2>
<ol>
<li><strong>Design for the smallest size first.</strong> If it works as a 16-pixel favicon, it will work everywhere else.</li>
<li><strong>Fewer weights, fewer colours.</strong> Every extra one is something that can break when the logo is shrunk, inverted or printed.</li>
<li><strong>Let one detail carry the personality.</strong> Here it&#39;s a single dot. Everything else stays out of its way.</li>
<li><strong>Trust your eyes over the grid.</strong> Overshoot, optical spacing and an off-centre icon all look wrong on paper and right on screen.</li>
</ol>
]]></content:encoded>
    </item>
    <item>
      <title>Building AutoShare: On-Device Face Recognition for Family Albums, and an API Apple Doesn’t Have</title>
      <link>https://www.sherlabs.com/blog/building-autoshare</link>
      <guid isPermaLink="true">https://www.sherlabs.com/blog/building-autoshare</guid>
      <pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate>
      <dc:creator>Nilesh Jayanandana</dc:creator>
      <description>AutoShare finds new photos of the people you choose and gets them into the right album, with face recognition that never leaves the iPhone. How it works, the Shared Album wall I hit, and the bugs that taught me the most.</description>
      <enclosure url="https://www.sherlabs.com/blog/building-autoshare/cover.jpg" type="image/jpeg" length="0" />
      <content:encoded><![CDATA[<img src="https://www.sherlabs.com/blog/building-autoshare/cover.jpg" alt="AutoShare: photos of the people you choose, ready for the album they belong in" /><figure><img src="https://www.sherlabs.com/blog/building-autoshare/01-ready-for-grandma.jpg" alt="AutoShare gets new photos of Mia ready for the grandparents' album" width="660" height="1434" class="post-shot" loading="lazy" decoding="async" /><figcaption>AutoShare gets new photos of Mia ready for the grandparents' album</figcaption></figure>
<p>Every family has the same request. The grandparents want every photo of the kid. Not the best ten, every one. My son&#39;s grandparents are no different, and for a long time the system was me, once a week, scrolling back through the camera roll, picking out the photos with him in them and adding them to a shared album.</p>
<p>Apple Photos already knows who is in your photos. It has a People album that recognises faces on the device. So I assumed an app could ask &quot;give me the new photos of this person&quot; and put them in an album.</p>
<p>It can&#39;t. Apple doesn&#39;t share People data with apps. That one fact shaped the whole product, and it led to a second, harder one I&#39;ll get to below.</p>
<p><strong>AutoShare</strong> is what I built instead. You pick a face once, choose an album, and AutoShare finds new photos of that person in Apple Photos and gets them into it. The face recognition runs entirely on the iPhone. It&#39;s free and live on the App Store now. This is how it works, the wall I hit, and the bugs that taught me the most.</p>
<h2>The Rules I Set First</h2>
<p>I made the same call I made with Snipo: no server. With AutoShare it wasn&#39;t really a choice. These are photos of children, and a face print is about as personal as data gets. I didn&#39;t want any of it leaving the phone, and I didn&#39;t want to be the person holding it.</p>
<p>So the rules were:</p>
<ul>
<li><strong>Everything runs on the iPhone.</strong> Face detection, recognition, matching, the lot.</li>
<li><strong>No account, no backend, no analytics, no ads, no third-party code.</strong></li>
<li><strong>Never claim something happened when it didn&#39;t.</strong> This one turned out to matter more than I expected.</li>
</ul>
<h2>The Stack</h2>
<p>AutoShare is a native iOS app, and I leaned on Apple&#39;s frameworks for almost everything:</p>
<ul>
<li><strong>Swift 6</strong> with strict concurrency, and <strong>SwiftUI</strong> for every screen</li>
<li><strong>SwiftData</strong> for storage, behind a single store actor that every write goes through</li>
<li><strong>PhotoKit</strong> to read the library and add photos to albums</li>
<li><strong>Vision</strong> to find faces, and <strong>Core ML</strong> to recognise them</li>
<li><strong>BackgroundTasks</strong> to ask iOS for scanning time, and <strong>App Intents</strong> for Shortcuts</li>
<li><strong>UserNotifications</strong> to say when photos are ready to share</li>
</ul>
<p>There&#39;s no backend at all. The project is generated with XcodeGen. The first commit was on November 2, the first App Store build went up within days, and version 1.0.1 is 53 commits in.</p>
<h2>Teaching It a Face</h2>
<p>Apple&#39;s Vision framework is very good at finding faces. It gives you a box around each one, landmarks for the eyes, nose and mouth, and a quality score. What it doesn&#39;t give you is identity. It can tell you there&#39;s a face; it can&#39;t tell you whose.</p>
<p>For that I bundled a face-recognition model: FaceNet (Inception-ResNet-v1), converted to Core ML. It takes a 160 × 160 face crop and returns 512 numbers, an embedding, where photos of the same person land close together and different people land far apart. The model is about 45 MB. I checked the Core ML conversion against the original PyTorch model and the outputs agree to a cosine similarity of 0.99997. I deliberately didn&#39;t use the popular ArcFace weights, because their licence is non-commercial.</p>
<p>Getting good embeddings is mostly about the crop. The model was trained on faces that were upright and centred with a margin around them, so AutoShare does the same: find the face with Vision, straighten it using the eye landmarks, take a square about 1.3 times the size of the face, and resize it.</p>
<figure><img src="https://www.sherlabs.com/blog/building-autoshare/pipeline.svg" alt="The AutoShare pipeline" width="1400" height="820" loading="lazy" decoding="async" /><figcaption>The AutoShare pipeline</figcaption></figure>
<p>Then matching is just maths. You teach AutoShare a person by picking a few of their faces. Every new face gets compared with those reference faces, and if it&#39;s close enough, the photo matches. Embeddings are stored as 16-bit floats, so a face takes 1 KB, and comparisons run on Apple&#39;s Accelerate framework.</p>
<figure><img src="https://www.sherlabs.com/blog/building-autoshare/02-teach-a-face.jpg" alt="Teach a face once" width="660" height="1434" class="post-shot" loading="lazy" decoding="async" /><figcaption>Teach a face once</figcaption></figure>
<p>Picking faces by hand is a chore, so AutoShare also does <strong>Face Discovery</strong>. As it scans, it groups faces that look like the same person and shows them under <em>Suggested</em> on the People tab. You name a group, and that person is ready. Groups need faces from at least three different photos before they&#39;re shown, so a few frames from one video can&#39;t fake a person.</p>
<p>When AutoShare isn&#39;t sure about a match, it doesn&#39;t guess. It asks: &quot;Is this Mia?&quot; Tap <strong>Not them</strong> and it learns that face isn&#39;t her, and you can undo that if you tap the wrong one.</p>
<figure><img src="https://www.sherlabs.com/blog/building-autoshare/04-people.jpg" alt="The People tab, with a suggested face to name" width="660" height="1434" class="post-shot" loading="lazy" decoding="async" /><figcaption>The People tab, with a suggested face to name</figcaption></figure>
<h2>The Wall: Shared Albums</h2>
<p>Here&#39;s the problem I hadn&#39;t seen coming.</p>
<p>The album the grandparents actually look at is an <strong>iCloud Shared Album</strong>. And iOS has no public API for an app to post to one. PhotoKit can add photos to your own albums with no interaction at all. It cannot touch a Shared Album. I checked the iOS 27 SDK headers to be sure.</p>
<p>What iOS 27 does give you is Apple&#39;s own posting sheet: an app can open it with the photos already selected, and the user taps <strong>Post</strong>. That&#39;s it.</p>
<p>I had a few ways out, and most of them were bad:</p>
<ul>
<li><strong>Fake it.</strong> UI scripting, simulated taps, anything that pretends to be the user. That&#39;s against the rules, it would break with every iOS update, and it&#39;s exactly the kind of thing that should make people distrust an app that has their photos.</li>
<li><strong>Say it&#39;s automatic and hope nobody checks.</strong> No.</li>
<li><strong>Shortcuts.</strong> I built App Intents that hand pending photos to Shortcuts and mark a batch as shared afterwards. Whether a Shortcuts action can post to a Shared Album without its own confirmation is something I couldn&#39;t verify, so it&#39;s there as an experiment and labelled that way, not as the main path.</li>
</ul>
<p>So I went with the honest version. Regular Apple Photos albums are fully automatic. For Shared Albums, AutoShare does all the finding and batching, sends a notification when a batch is ready, and opens Apple&#39;s sheet with the photos in it. You tap Post. One tap per batch, and the copy says so everywhere: onboarding, the rule screen and Settings.</p>
<figure><img src="https://www.sherlabs.com/blog/building-autoshare/03-you-decide.jpg" alt="Nothing goes out without you" width="660" height="1434" class="post-shot" loading="lazy" decoding="async" /><figcaption>Nothing goes out without you</figcaption></figure>
<p>There&#39;s a rule in the code to match. A photo is only marked as shared when Apple&#39;s sheet reports it was posted. Until then it&#39;s &quot;ready to share&quot;, never &quot;sent&quot;.</p>
<h2>Bugs That Taught Me Something</h2>
<p>Most of the interesting work happened after it ran on my own phone, against my own library.</p>
<h3>&quot;Is this my son?&quot; next to his mum&#39;s face</h3>
<p>Review asked &quot;Is this <em>my son</em>?&quot; and showed a photo of his mum holding him. Looking at the whole photo, the answer was clearly yes. He was right there. But the match was on <em>her</em> face. AutoShare had matched the wrong face in a group photo, and the question showed the photo, not the face it matched.</p>
<p>If I said no, the photo disappeared, even though he was in it. If I said yes, AutoShare learned that her face was his. Both were wrong.</p>
<p>The fix was two parts. Review now shows a crop of the actual face that matched, so you&#39;re answering about the right face. And when you say no on a photo with other faces in it, AutoShare asks a follow-up: &quot;Is he elsewhere in this photo?&quot; Tap the right face and the photo goes through with the correct match. Tap &quot;No&quot; and it&#39;s removed as before.</p>
<h3>The total that kept climbing</h3>
<p>The scan card said &quot;48 of 416&quot;, and every time it got close to the end, the end moved. Scanning works through history in pages, and new pages get added as the queue drains, so the total really was growing. It just looked like the app was lying.</p>
<p>Now the card only shows &quot;N of M&quot; once the total is final. Until then it says &quot;N photos checked&quot;, which is true the whole time.</p>
<h3>Activity that buried the important line</h3>
<p>The Activity tab logged &quot;N new photos ready to share&quot; once per batch. On a big first scan, that was over 500 entries, which pushed the one line I cared about, &quot;Shared&quot;, out of the list. Now there&#39;s one entry per scan, empty scans aren&#39;t logged, and sharing entries are always shown.</p>
<h3>2,500 photos that never arrived</h3>
<p>This was the worst one. After a big post to the grandparents&#39; album, about 2,500 photos simply weren&#39;t there. Apple&#39;s sheet had reported success. AutoShare had marked them as shared. They weren&#39;t in the album.</p>
<p>The cause was Apple&#39;s Shared Album limits. Shared albums created before iOS 27 hold at most 5,000 photos and videos, and accept 1,000 uploads per person per hour. Albums created on iOS 27 or later don&#39;t have the per-album cap. I&#39;d actually had a 5,000 check in the app and removed it shortly before, after reading that iOS 27 had dropped the cap. It has, but only for albums created on iOS 27 or later. Ours was older.</p>
<figure><img src="https://www.sherlabs.com/blog/building-autoshare/shared-album-parts.svg" alt="Splitting a big post to fit Apple's limits" width="1400" height="640" loading="lazy" decoding="async" /><figcaption>Splitting a big post to fit Apple's limits</figcaption></figure>
<p>The fix:</p>
<ul>
<li><strong>Split big posts into parts.</strong> Each batch holds at most what Apple will accept right now, counting the last hour and the last day across all rules, oldest photos first.</li>
<li><strong>Notify when the next part can go.</strong> If the hour is used up, AutoShare schedules a notification for when it frees up. Tapping it opens Apple&#39;s sheet for the next part. You still tap Post every time.</li>
<li><strong>Warn before an album fills up.</strong> If an older album is getting close to 5,000, the rule screen says so and suggests a new Shared Album.</li>
</ul>
<p>The lesson for me: when you can&#39;t verify the result, trust it less. AutoShare can&#39;t read a Shared Album&#39;s contents, so it has to respect the limits up front instead of finding out later.</p>
<h2>Background Scanning, Honestly</h2>
<p>There&#39;s no &quot;a photo was added&quot; trigger for apps on iOS. AutoShare scans when you open it and asks iOS for background time, and iOS decides when, or whether, that happens. The app says so rather than promising &quot;instant&quot;. In Settings you can choose Automatic, Battery-conscious, Wi-Fi preferred or Manual only.</p>
<p>Performance mattered more than I expected, too. An early build had 13-second hangs on device, and the trace showed store work running on the main thread. SwiftData&#39;s <code>@ModelActor</code> ran jobs on whichever thread called it, which was often the UI. I rewrote the store to conform to <code>ModelActor</code> by hand, with its own serial queue, so store work never runs on main.</p>
<h2>Clean Up</h2>
<p>Once AutoShare was looking at every photo anyway, a second job was easy to add: finding the clutter. The Utilities tab finds old screenshots, photos of documents and receipts, and blurry shots. Blurry means two signals agree: a low sharpness score and a low Vision aesthetics score.</p>
<p>It never picks favourites or rated photos for you, and it can&#39;t delete anything by itself. iOS asks you to confirm, and deleted photos go to Recently Deleted like any other.</p>
<figure><img src="https://www.sherlabs.com/blog/building-autoshare/07-clean-up.jpg" alt="Clean Up" width="660" height="1434" class="post-shot" loading="lazy" decoding="async" /><figcaption>Clean Up</figcaption></figure>
<h2>Privacy, in Practice</h2>
<p>&quot;Private by design&quot; is easy to say, so here&#39;s what it means in AutoShare:</p>
<ul>
<li><strong>Face data never leaves the phone.</strong> There&#39;s no server to send it to.</li>
<li><strong>Protected storage.</strong> Face data lives in the app&#39;s own store, protected by iOS data protection.</li>
<li><strong>No photo identifiers in logs.</strong> Diagnostics never include which photos or faces were involved.</li>
<li><strong>Delete everything in one tap.</strong> Settings › Delete All Face Data removes every reference face and match.</li>
<li><strong>Only people who are happy to be included.</strong> Please only add people who are fine with it, or children you&#39;re the parent or guardian of.</li>
</ul>
<figure><img src="https://www.sherlabs.com/blog/building-autoshare/06-private.jpg" alt="Private by design" width="660" height="1434" class="post-shot" loading="lazy" decoding="async" /><figcaption>Private by design</figcaption></figure>
<h2>Shipping It</h2>
<p>The first App Store screenshots used cartoon demo people. They were fine, but they didn&#39;t look like a photo app. Nobody&#39;s family album is full of cartoons.</p>
<p>So I added a demo mode, DEBUG builds only, that fills the app with stock photos from Pexels instead. Mia, Leo and Emma in the screenshots are stock photos, not real users and not my family. Each photo comes with the real face rectangles found on the device, so group photos crop to the right face, just like in the real app. None of those photos ship in the app. I captured the screens in the simulator and framed them in a browser.</p>
<figure><img src="https://www.sherlabs.com/blog/building-autoshare/05-every-photo.jpg" alt="Every photo of them, in one place" width="660" height="1434" class="post-shot" loading="lazy" decoding="async" /><figcaption>Every photo of them, in one place</figcaption></figure>
<p>The promo video was rendered frame by frame in headless Chrome, with a soundtrack I synthesised rather than licensed:</p>
<p><video src="https://www.sherlabs.com/autoshare/autoshare-promo.mp4" poster="https://www.sherlabs.com/autoshare/autoshare-promo.jpg" width="720" height="1280" controls muted playsinline preload="metadata"></video></p>
<p>Version 1.0.0 is live, and 1.0.1, with the new screenshots and the Shared Album fixes above, is in review as I write this.</p>
<h2>What&#39;s Next</h2>
<ul>
<li>Seeing how the face matching holds up on many more libraries than mine.</li>
<li>Fewer taps for Shared Albums, if Apple ever adds a way. I&#39;ll keep checking each iOS release.</li>
<li>And, as with Snipo, keeping it small and private.</li>
</ul>
<p>If someone in your family wants every photo of the kids, give it a try:</p>
<p><strong><a href="https://apps.apple.com/app/id6830419182" target="_blank" rel="noopener noreferrer">Download AutoShare on the App Store</a></strong></p>
<p>AutoShare is free, has no ads and no account, and it&#39;s built by one person in Australia. More about it is on the <a href="https://www.sherlabs.com/autoshare">AutoShare page</a>. If it&#39;s useful to you, you can <a href="https://www.sherlabs.com/support-us">support it here</a>.</p>
]]></content:encoded>
    </item>
    <item>
      <title>Building Snipo: A Local-First AI App for the Reels You Save and Never Open Again</title>
      <link>https://www.sherlabs.com/blog/building-snipo</link>
      <guid isPermaLink="true">https://www.sherlabs.com/blog/building-snipo</guid>
      <pubDate>Wed, 04 Nov 2026 00:00:00 GMT</pubDate>
      <dc:creator>Nilesh Jayanandana</dc:creator>
      <description>Snipo reads the reels you save, files them, and lets you ask your whole library anything, with no server and AI that runs on your iPhone or your own ChatGPT plan. The architecture, the decisions, and what broke along the way.</description>
      <enclosure url="https://www.sherlabs.com/blog/building-snipo/promo.jpg" type="image/jpeg" length="0" />
      <content:encoded><![CDATA[<img src="https://www.sherlabs.com/blog/building-snipo/promo.jpg" alt="Snipo: your saved reels, finally useful" /><figure><img src="https://www.sherlabs.com/blog/building-snipo/promo.jpg" alt="Snipo: your saved reels, finally useful" width="1400" height="787" loading="lazy" decoding="async" /><figcaption>Snipo: your saved reels, finally useful</figcaption></figure>
<p>We all do it. You&#39;re scrolling, you see a reel that&#39;s actually useful. A pasta recipe, a 10-minute workout, a café in Tokyo you swear you&#39;ll visit. You tap save.</p>
<p>And you never open it again.</p>
<p>My Instagram saved folder had hundreds of these. TikTok had more. YouTube had a &quot;Watch later&quot; playlist that was really a &quot;Watch never&quot; playlist. All that know-how, sitting in three different apps, completely useless.</p>
<p>So I built <strong>Snipo</strong>. You share a reel to it, it reads the reel for you, files it, and lets you ask your whole library anything. &quot;What do I need for that garlic pasta?&quot; &quot;Plan me a weekend in Lisbon from my saves.&quot; Then it turns the answer into a checklist you can actually tick off.</p>
<p>It&#39;s live on the App Store today, and this is how I built it: the architecture, the decisions, and the things that broke along the way.</p>
<h2>The Big Decision: No Server</h2>
<p>Snipo didn&#39;t start local-first. The first version, shipped in February, ran on Supabase. It had Google sign-in, ads and social features. It worked, but it felt wrong.</p>
<p>Think about what&#39;s in someone&#39;s saved reels. Recipes, sure. But also workouts, health tips, travel plans, things they want to buy. That&#39;s a pretty personal picture of someone&#39;s life. I didn&#39;t want it sitting on my server, and honestly, I didn&#39;t want to be paying to host it either.</p>
<p>So in October I tore it out. One commit touched 524 files: +41.6k lines, −43.2k lines. No Supabase. No accounts. No ads. No analytics.</p>
<p>The rule became simple:</p>
<ul>
<li><strong>Everything lives on the phone.</strong> Library, chats, plans, settings.</li>
<li><strong>No Snipo server, ever.</strong> There is literally nothing for me to leak.</li>
<li><strong>AI either runs on the iPhone, or on the user&#39;s own ChatGPT plan.</strong> Not mine.</li>
</ul>
<p>That one decision shaped everything else.</p>
<figure><img src="https://www.sherlabs.com/blog/building-snipo/architecture.png" alt="Snipo system overview" width="1400" height="980" loading="lazy" decoding="async" /><figcaption>Snipo system overview</figcaption></figure>
<h2>The Stack</h2>
<p>Nothing exotic here. I&#39;ve been using Flutter since 2018 (I wrote a &quot;first impressions&quot; post back then), and it&#39;s matured into a genuinely great way to ship mobile apps.</p>
<ul>
<li><strong>Flutter 3.41</strong> with <strong>Riverpod</strong> for state and <strong>go_router</strong> for navigation</li>
<li><strong>drift / SQLite</strong> as the on-device database</li>
<li><strong>Swift</strong> for everything iOS-specific: the Share Extension, Siri intents, Vision OCR, on-device speech, and Apple&#39;s Foundation Models, all talking to Dart over method channels</li>
<li><strong>iCloud Drive</strong> for backups: a SQLite snapshot, daily, newest five kept, never including tokens</li>
</ul>
<p>Today the app is about 59,000 lines of Dart and 1,350 lines of Swift, with 679 tests.</p>
<h2>Saving a Reel: Harder Than It Looks</h2>
<p>The core flow sounds trivial. User taps Share, picks Snipo, done. In practice, this is where most of the engineering went.</p>
<figure><img src="https://www.sherlabs.com/blog/building-snipo/save-pipeline.png" alt="The save pipeline" width="1400" height="1010" loading="lazy" decoding="async" /><figcaption>The save pipeline</figcaption></figure>
<p><strong>1. The Share Extension does almost nothing.</strong> iOS share extensions are memory-limited and get killed easily. So mine doesn&#39;t touch the network. It writes the link into a shared App Group and opens the main app through a URL scheme. That&#39;s it.</p>
<p><strong>2. Save first, then work.</strong> The link goes into the database <em>before</em> any network call. Then every step is checkpointed: pending → scraped → captured → done. If iOS suspends the app halfway through (and it will), the save resumes where it stopped instead of starting again.</p>
<p><strong>3. Scraping without API keys.</strong> There&#39;s no official &quot;give me this reel&#39;s content&quot; API for any of these platforms. So each platform gets its own scraper. TikTok needs an iPhone user agent to get past its firewall. YouTube goes through its internal mobile clients. Instagram, Facebook and LinkedIn each have their own quirks. Captions, OCR on frames (Vision) and transcripts (on-device speech) all run on the phone.</p>
<p><strong>4. Read it, then file it.</strong> The AI writes a short summary, pulls out the steps, places, products or recipes, and picks the right collection. The user sees it happen live: <em>Fetching the link → Listening → Reading it → Finding its home.</em></p>
<figure><img src="https://www.sherlabs.com/blog/building-snipo/screenshot_04_save.jpg" alt="Saving a reel in Snipo" width="700" height="1521" class="post-shot" loading="lazy" decoding="async" /><figcaption>Saving a reel in Snipo</figcaption></figure>
<h2>Bring Your Own AI</h2>
<p>Here&#39;s the problem with &quot;no server&quot; and &quot;AI app&quot; in the same sentence: someone has to pay for the AI.</p>
<p>My answer was <strong>Sign in with ChatGPT</strong>. Users connect the ChatGPT Plus or Pro plan they already pay for. It&#39;s OAuth with PKCE through a loopback server on the phone. Tokens live in the iOS Keychain, and requests go straight from the phone to OpenAI with <code>store: false</code>. I never see any of it.</p>
<p>That worked, but it left out everyone without ChatGPT. So version 2.0.2 (rolling out now) adds <strong>on-device AI</strong>:</p>
<ul>
<li><strong>Apple Foundation Models</strong> on iPhones with Apple Intelligence (iOS 26+). The best part is guided generation: instead of asking the model for JSON and praying, I define typed Swift structs with <code>@Generable</code>, and the model fills them in directly. Six structured outputs (item analysis, collection assignment, three kinds of summaries, cover design) all go through this. No more broken JSON.</li>
<li><strong>Gemma 4 E2B</strong> for iPhones with 6 GB of RAM that don&#39;t have Apple Intelligence. It&#39;s a 2.6 GB download, only after the user confirms, resumable, and deletable from Settings.</li>
</ul>
<figure><img src="https://www.sherlabs.com/blog/building-snipo/ai-router.png" alt="How Snipo picks a model" width="1400" height="1000" loading="lazy" decoding="async" /><figcaption>How Snipo picks a model</figcaption></figure>
<p>One rule I was strict about: if the user chose on-device AI and it isn&#39;t available, Snipo shows an error. It <strong>never</strong> quietly falls back to sending their content to ChatGPT. If you say &quot;keep it on my phone,&quot; it stays on your phone.</p>
<h2>Ask: RAG Without Embeddings</h2>
<p>My favourite feature is Ask. You type a question, and Snipo answers from your own reels, with source chips you can tap.</p>
<figure><img src="https://www.sherlabs.com/blog/building-snipo/ask-flow.png" alt="Ask flow" width="1400" height="960" loading="lazy" decoding="async" /><figcaption>Ask flow</figcaption></figure>
<p>I didn&#39;t use embeddings. For a personal library of a few hundred reels, a good keyword ranker with field weights does the job and costs nothing. Long YouTube transcripts get split into BM25 chunks with approximate timestamps, so an answer can point to roughly where in the video something was said.</p>
<p>The interesting constraint is context size. ChatGPT gets about 12,000 characters of library context. Apple&#39;s on-device model gets about 4,000. So retrieval has to be sharp: the best few reels, the best few chunks, nothing wasted.</p>
<figure><img src="https://www.sherlabs.com/blog/building-snipo/screenshot_02_ask.jpg" alt="Asking your library" width="700" height="1521" class="post-shot" loading="lazy" decoding="async" /><figcaption>Asking your library</figcaption></figure>
<h2>Plans That Remember</h2>
<p>An answer is nice. A checklist is useful.</p>
<p>Any answer can be saved as a plan. Tick things off as you go. And here&#39;s the small detail that makes it feel smart: everything you&#39;ve ticked is fed back into the next conversation as &quot;already done.&quot; Ask for a new weekend plan and it skips the café you already went to.</p>
<figure><img src="https://www.sherlabs.com/blog/building-snipo/screenshot_03_plans.jpg" alt="Plans you tick off" width="700" height="1521" class="post-shot" loading="lazy" decoding="async" /><figcaption>Plans you tick off</figcaption></figure>
<h2>Things That Broke</h2>
<p>No build story is honest without these.</p>
<ul>
<li><strong>Cold-start shares showed &quot;Page Not Found&quot;.</strong> Sharing to Snipo while it wasn&#39;t running made Flutter treat the share URL as a deep link to a route that didn&#39;t exist. Fix: turn off Flutter&#39;s deep linking and add a guard that swallows the share URL.</li>
<li><strong>A launch crash on iOS 27.</strong> Fixed by moving to the scene-based app lifecycle.</li>
<li><strong>Retries deleting reels.</strong> Two retry passes running at once decided a reel was its own duplicate and deleted it. Fix: per-reel claims with a timeout, which I later raised from 3 to 8 minutes.</li>
<li><strong>&quot;Paused&quot; isn&#39;t &quot;dropped&quot;.</strong> When iOS suspends the app, it kills the network connection, not the link. I had to split &quot;Reading paused&quot; from &quot;Connection dropped&quot; so users weren&#39;t told something failed when it was just waiting.</li>
<li><strong>Six visual redesigns in three days.</strong> Then a proper readability pass: 4.5:1 contrast everywhere and Dynamic Type tests, because a pretty app nobody can read is useless.</li>
</ul>
<h2>Building With AI as a Teammate</h2>
<p>I wrote last year about the rise of the Product Engineer: someone who doesn&#39;t just write code, but owns the product, the UX and the decisions. Snipo is me trying to live that.</p>
<p>I built it with Claude Code as a pair programmer. Of 216 commits, 95 have Claude as co-author. It wrote a lot of code, but the more interesting part is what it freed me up to do: obsess over the share flow, redo the design six times, and make the calls about privacy that actually define the product.</p>
<p>AI didn&#39;t replace the engineering. It changed which parts of it I spend my time on.</p>
<h2>What&#39;s Next</h2>
<ul>
<li><strong>Android</strong> is coming soon.</li>
<li>More on-device smarts as Apple&#39;s models get bigger context windows.</li>
<li>And as always, fewer taps between &quot;I saw something useful&quot; and &quot;I actually did it.&quot;</li>
</ul>
<p>If you&#39;ve got a saved folder full of reels you&#39;ll never open again, give it a try:</p>
<p><strong><a href="https://apps.apple.com/app/id6759781527" target="_blank" rel="noopener noreferrer">Download Snipo on the App Store</a></strong></p>
<p>Snipo is free, has no ads and no account, and it&#39;s built by one person. If it&#39;s useful to you, you can <a href="https://www.sherlabs.com/support-us">support it here</a>. And I&#39;d genuinely love to hear what you save first.</p>
]]></content:encoded>
    </item>
  </channel>
</rss>
