<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Shivam Anil Jha</title>
    <description>The latest articles on DEV Community by Shivam Anil Jha (@shivamj56).</description>
    <link>https://dev.to/shivamj56</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4247104%2F827ebe89-4a3b-4e93-bb19-da2c0d0f8056.jpg</url>
      <title>DEV Community: Shivam Anil Jha</title>
      <link>https://dev.to/shivamj56</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/shivamj56"/>
    <language>en</language>
    <item>
      <title>A calmer family TV, without the risk: the safety rules behind Orbit OS</title>
      <dc:creator>Shivam Anil Jha</dc:creator>
      <pubDate>Tue, 17 Nov 2026 04:32:09 +0000</pubDate>
      <link>https://dev.to/shivamj56/a-calmer-family-tv-without-the-risk-the-safety-rules-behind-orbit-os-4g15</link>
      <guid>https://dev.to/shivamj56/a-calmer-family-tv-without-the-risk-the-safety-rules-behind-orbit-os-4g15</guid>
      <description>&lt;p&gt;I have already written about &lt;a href="https://shivamjha.dev/orbit-os.html" rel="noopener noreferrer"&gt;how Orbit OS works&lt;/a&gt;, and here on DEV about &lt;a href="https://dev.to/shivamj56/i-built-a-phone-app-that-debloats-google-tv-over-wi-fi-with-an-adb-client-in-pure-dart-35lk"&gt;its ADB client in pure Dart&lt;/a&gt;. This post is about who it is for, and the rules I wrote before any feature: a tool that changes a family's TV has to be harder to break than the TV itself.&lt;/p&gt;

&lt;p&gt;Orbit OS is a free Android app that sets up a Google TV or Android TV from a phone, and undoes it. It is open source under GPL-3.0.&lt;/p&gt;

&lt;h2&gt;
  
  
  The TV everyone shares
&lt;/h2&gt;

&lt;p&gt;Orbit was built and tested on the TV in our family room, a Xiaomi TV A Pro with Google TV 14. A new Google TV starts with a home screen full of rows nobody asked for, and dozens of preinstalled apps that wake up with the TV, and this one was no different. For someone who only wants the news, a serial or a bhajan, that is a lot of screen to get through.&lt;/p&gt;

&lt;p&gt;The people who use a family TV most are rarely the ones who would type an ADB command. That shaped what Orbit puts on it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Orbit Home&lt;/strong&gt; shows only the apps the family uses, with a greeting, the date and the weather, and colours that change for festivals like Diwali, Holi and Chhath.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Orbit Live TV&lt;/strong&gt; has about 30 free Hindi channels (news, serials, bhakti, kids and movies), played from the broadcasters' own streams.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The phone is the remote&lt;/strong&gt;, and in Help mode you can see the TV's screen on your phone and press its buttons for someone at home, even when you are away, after a Google sign-in. The TV shows who is watching, and help ends by itself.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;But the person who runs Orbit's setup is usually not the person who lives with the TV every day. If setup goes wrong, someone's parents are left with a TV that boots to an empty screen. So the safety rules came first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 1: write the TV down before touching it
&lt;/h2&gt;

&lt;p&gt;The first step of setup changes nothing. It records how the TV is right now: which app is the home screen, which accessibility services are on, and the screensaver. Every app that setup then switches off is added to that record the moment Android confirms the change, and each line is also sent to the phone, which keeps its own copy.&lt;/p&gt;

&lt;p&gt;If the TV ever loses its record, the phone sends its copy back before undo runs. And if setup runs a second time, it keeps the first record, because that is the one that knows how the TV was.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 2: switch off, never delete
&lt;/h2&gt;

&lt;p&gt;Orbit switches apps off with Android's own disable command. It never uninstalls an app that came with the TV; the only apps it ever removes are its own. A list of core packages is never touched at all: the system UI, settings, the permission controller, the package installer, Android's providers and the TV's own tuner services. On a TV it has not been tested on, Orbit offers only the changes known to be safe on any TV.&lt;/p&gt;

&lt;p&gt;On the test TV, setup switched off 32 preinstalled apps and deleted none.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 3: the home screen goes last
&lt;/h2&gt;

&lt;p&gt;The stock home screen is the one thing whose loss leaves a TV with nothing to show. So the order is fixed. Orbit installs Orbit Home, starts it, and waits for Android to report that it really came up. Only then does it switch off the stock launcher and the packages tied to it. Then it makes Orbit Home the home screen and asks Android what the home screen is now.&lt;/p&gt;

&lt;p&gt;If the answer is not Orbit Home, setup switches the stock packages back on, makes the old home screen the home again, and tells you the TV was put back the way it was. If Orbit Home fails to install or start, setup stops right there and the TV keeps its own home screen.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 4: undo in the safest order
&lt;/h2&gt;

&lt;p&gt;Undo reads the record and works through it by risk. The TV's own home screen comes back first, so there is never a moment without one. Then the apps are switched back on and the settings restored. Orbit's own apps go last, and never the one that is still the home screen.&lt;/p&gt;

&lt;h2&gt;
  
  
  Smaller things that follow from the same idea
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Setup asks five plain questions and shows the whole plan on the phone before it changes anything.&lt;/li&gt;
&lt;li&gt;The TV's original setup can be backed up to the family's own Google Drive.&lt;/li&gt;
&lt;li&gt;The pairing token that connects the phone to the TV is deleted from the TV once it has been handed over, along with the install files.&lt;/li&gt;
&lt;li&gt;The scripts on the TV are written for Android's own small shell (mksh and toybox), not bash, because that is what a TV has.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;On the test TV, setup takes about 5 minutes from pressing Allow to the finished home screen. The &lt;a href="https://shivamjha.dev/orbit-os.html" rel="noopener noreferrer"&gt;case study&lt;/a&gt; has the rest of the measured numbers.&lt;/p&gt;

&lt;p&gt;Orbit OS is free on &lt;a href="https://play.google.com/store/apps/details?id=dev.shiv.orbit&amp;amp;referrer=utm_source%3Ddev.to%26utm_medium%3Dblog"&gt;Google Play&lt;/a&gt;, and the code is on &lt;a href="https://github.com/shivamj56/orbit-os" rel="noopener noreferrer"&gt;GitHub&lt;/a&gt;.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This post first appeared on my blog at &lt;a href="https://shivamjha.dev/blog/orbit-os-safe-setup-for-family-tvs.html" rel="noopener noreferrer"&gt;shivamjha.dev&lt;/a&gt;. I'm Shivam Anil Jha, an AI architect and full-stack developer from Mumbai. I also build &lt;a href="https://nexera-ai-xi.vercel.app/" rel="noopener noreferrer"&gt;NexeraAI&lt;/a&gt; and &lt;a href="https://clonix-tau.vercel.app/" rel="noopener noreferrer"&gt;Clonix&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>android</category>
      <category>opensource</category>
      <category>showdev</category>
      <category>ux</category>
    </item>
    <item>
      <title>How Clonix turns a website into design tokens for Cursor, Claude Code and Windsurf</title>
      <dc:creator>Shivam Anil Jha</dc:creator>
      <pubDate>Tue, 17 Nov 2026 04:28:20 +0000</pubDate>
      <link>https://dev.to/shivamj56/how-clonix-turns-a-website-into-design-tokens-for-cursor-claude-code-and-windsurf-5aie</link>
      <guid>https://dev.to/shivamj56/how-clonix-turns-a-website-into-design-tokens-for-cursor-claude-code-and-windsurf-5aie</guid>
      <description>&lt;p&gt;Ask an AI code editor to make something look like a site you like, and you get something close but wrong. The blue is a shade off, the headings use Tailwind's default sizes, and the shadows are invented. The editor never saw the site; it is working from memory. Clonix is my answer to that: it measures the real page in your browser and hands your editor the numbers.&lt;/p&gt;

&lt;p&gt;Clonix is a Chrome extension with a small backend, and I built it with the Sqaron team. This post is about the half that reads a page: how a live website becomes design tokens and editor rules.&lt;/p&gt;

&lt;h2&gt;
  
  
  Measuring what a visitor actually sees
&lt;/h2&gt;

&lt;p&gt;The scan is one function that the extension injects into the page's main world, so it can read computed styles, the loaded fonts and the live DOM, not just the HTML source. It makes a single walk over the elements a visitor can actually see. Anything that is &lt;code&gt;display: none&lt;/code&gt;, invisible, fully transparent, parked off-screen at -9999px or clipped to a 1px box (skip links and screen-reader text) doesn't count. Decorations marked &lt;code&gt;aria-hidden&lt;/code&gt; still count for what they draw, but never as buttons.&lt;/p&gt;

&lt;p&gt;The walk has a budget: at most 6,000 visible elements in 300 ms. Whatever it doesn't reach is counted and reported with the scan, never guessed.&lt;/p&gt;

&lt;p&gt;Code reads CSS, but people see paint. Most of the rules in the walk come from that difference:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Colours are weighted by how much of the screen they paint and how many characters are set in them, not by how often they appear in a stylesheet.&lt;/li&gt;
&lt;li&gt;A fully transparent colour paints nothing, so it is ignored. An early version turned &lt;code&gt;rgba(0, 0, 0, 0)&lt;/code&gt; into &lt;code&gt;#000000&lt;/code&gt;, which is black.&lt;/li&gt;
&lt;li&gt;A semi-transparent colour is blended over whatever is under it.&lt;/li&gt;
&lt;li&gt;A gradient background has no &lt;code&gt;background-color&lt;/code&gt;, so its colour is the average of its stops along the line.&lt;/li&gt;
&lt;li&gt;Newer colour formats such as &lt;code&gt;oklch()&lt;/code&gt; and &lt;code&gt;lab()&lt;/code&gt; are read by painting them into a 1-pixel canvas, which understands every syntax the browser does.&lt;/li&gt;
&lt;li&gt;The page background is the deepest full-page box with an opaque background, not just &lt;code&gt;body&lt;/code&gt;. On many sites a wrapper inside the app's root element paints over everything.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  From measurements to roles
&lt;/h2&gt;

&lt;p&gt;A list of twenty colours doesn't help an editor much. It needs to know what each one is for, so Clonix gives them roles:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Primary&lt;/strong&gt;: the most common filled button whose label has at least 3:1 contrast. If the site names that colour in a CSS variable such as &lt;code&gt;--primary&lt;/code&gt; or &lt;code&gt;--color-primary&lt;/code&gt;, the site's own choice wins.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ink&lt;/strong&gt;: the strongest of the colours the headings are set in.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Muted&lt;/strong&gt;: a quieter grey that still reads at 3:1 against the page.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Link and border&lt;/strong&gt;: the colour links are drawn in, and the hairline colour that boxes use.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Type works the same way. The h1, h2 and h3 styles come from the headings. Body is the size most of the other text is set in, in its most-used style, and small is the most-used size below that, when there is enough of it. Spacing gets a base of 8 or 4 pixels when that number divides at least 80% of the paddings, gaps and margins it measured; otherwise Clonix lists the values in use.&lt;/p&gt;

&lt;h2&gt;
  
  
  Writing it where each editor looks
&lt;/h2&gt;

&lt;p&gt;From one scan, Clonix writes the same design brief in the place each tool reads it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;.cursor/rules/clonix-design.mdc&lt;/code&gt;, a Cursor project rule that is always on&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;CLAUDE.md&lt;/code&gt;, Claude Code's project memory&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;.windsurfrules&lt;/code&gt; for Windsurf&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;AGENTS.md&lt;/code&gt; for Codex and other agents, and &lt;code&gt;.github/copilot-instructions.md&lt;/code&gt; for Copilot&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;tailwind.config.ts&lt;/code&gt; for Tailwind v3, &lt;code&gt;clonix-theme.css&lt;/code&gt; for Tailwind v4's &lt;code&gt;@theme&lt;/code&gt;, plain &lt;code&gt;tokens.css&lt;/code&gt;, and &lt;code&gt;design-tokens.json&lt;/code&gt; in the W3C design tokens format&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This step uses no network and no AI. It is plain code over the scan, so the same scan always gives the same files.&lt;/p&gt;

&lt;p&gt;Two decisions in those files matter more than they look. First, measured values never take Tailwind's own scale names. A site's heading size becomes &lt;code&gt;text-h1&lt;/code&gt;, its card radius &lt;code&gt;rounded-card&lt;/code&gt;, its shadow &lt;code&gt;shadow-elevation-1&lt;/code&gt;. If Clonix wrote the site's 15px text as &lt;code&gt;text-sm&lt;/code&gt;, it would quietly change what &lt;code&gt;text-sm&lt;/code&gt; means in your project, and the editor would trust the wrong name.&lt;/p&gt;

&lt;p&gt;Second, the brief is written as rules, not suggestions. Use the tokens exactly and never approximate. Use only the site's breakpoints, in the direction the site uses them, mobile-first or desktop-first. For dark mode, switch with the site's own selector and use its dark values instead of inverting colours by hand. If the page was scanned with its dark theme showing, the brief says so, so nobody mistakes those colours for the light theme.&lt;/p&gt;

&lt;h2&gt;
  
  
  Components, with AI
&lt;/h2&gt;

&lt;p&gt;The one step that uses AI is turning a section into a component. You pick a section of the page, and Clonix sends its cleaned HTML and the tokens to Claude, which writes a React and Tailwind, Vue 3, or plain HTML and CSS component. The instructions say to use the tokens exactly, to treat the captured HTML as a picture of the layout rather than code to copy, and to drop framework noise like hashed class names, tracking attributes and extra wrapper divs.&lt;/p&gt;

&lt;p&gt;Clonix can also save a page as an offline ZIP that runs with one click, but that is a story for another post.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I learned
&lt;/h2&gt;

&lt;p&gt;The editor files are the easy part. The work is in measuring what the eye sees: a heading whose highlighted word is in another colour, a link that is styled as a button, a skip link that is technically on the page. Each of those became a rule in the walk, and each rule made the files the editor reads a little more honest.&lt;/p&gt;

&lt;p&gt;You can see Clonix at &lt;a href="https://clonix-tau.vercel.app/" rel="noopener noreferrer"&gt;clonix-tau.vercel.app&lt;/a&gt;.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This post first appeared on my blog at &lt;a href="https://shivamjha.dev/blog/clonix-design-tokens-for-ai-editors.html" rel="noopener noreferrer"&gt;shivamjha.dev&lt;/a&gt;. I'm Shivam Anil Jha, an AI architect and full-stack developer from Mumbai and a co-founder of &lt;a href="https://sqaron.dev/" rel="noopener noreferrer"&gt;Sqaron Technologies&lt;/a&gt;. I also build &lt;a href="https://nexera-ai-xi.vercel.app/" rel="noopener noreferrer"&gt;NexeraAI&lt;/a&gt; and &lt;a href="https://github.com/shivamj56/orbit-os" rel="noopener noreferrer"&gt;Orbit OS&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>css</category>
      <category>ai</category>
      <category>tailwindcss</category>
      <category>showdev</category>
    </item>
    <item>
      <title>How I built NexeraAI, an AI career platform built on real job listings</title>
      <dc:creator>Shivam Anil Jha</dc:creator>
      <pubDate>Tue, 17 Nov 2026 04:26:49 +0000</pubDate>
      <link>https://dev.to/shivamj56/how-i-built-nexeraai-an-ai-career-platform-built-on-real-job-listings-5g84</link>
      <guid>https://dev.to/shivamj56/how-i-built-nexeraai-an-ai-career-platform-built-on-real-job-listings-5g84</guid>
      <description>&lt;p&gt;NexeraAI is the career platform I have been building since June 2026. You upload your résumé, it reads it, scores it, and shows which open jobs fit you and which skills you are missing. Around that sit referrals, skill-gap lessons with quizzes, a learning roadmap, interview practice, cover letters, a résumé builder and a salary check. Employers can post jobs on it too.&lt;/p&gt;

&lt;p&gt;The AI gets the attention, but most of the work went into two plainer problems: getting real, clean job listings, and making sure the AI never fills a gap with something made up. This post is about both.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Stack at a glance:&lt;/strong&gt; Next.js and TypeScript, PostgreSQL with Prisma, Auth.js, Tailwind CSS, and one small AI client for Groq, Google Gemini and OpenAI. Live at &lt;a href="https://nexera-ai-xi.vercel.app/" rel="noopener noreferrer"&gt;nexera-ai-xi.vercel.app&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the jobs come from
&lt;/h2&gt;

&lt;p&gt;An AI career site without real jobs is just a chat window. So the first thing I built was an importer. It reads listings from public sources: RemoteOK, Arbeitnow, and the public job boards that companies run on Greenhouse, Lever and Ashby. Three more, Adzuna, JSearch and Jooble, switch themselves on when their API keys are set. Each source is a small module that returns the same raw shape, so adding a new one doesn't touch the others.&lt;/p&gt;

&lt;p&gt;Every source formats things its own way, so each listing goes through one normaliser before it is saved:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;HTML becomes plain text. Greenhouse sends its HTML encoded twice, so it gets decoded twice.&lt;/li&gt;
&lt;li&gt;Broken characters are repaired. Text decoded with the wrong character set shows "â€™" where an apostrophe should be, and a small table puts the right character back.&lt;/li&gt;
&lt;li&gt;Remote, hybrid or on-site is read from the location and the title, and the level (entry, mid, senior or lead) from words like "Sr." or "Principal" in the title.&lt;/li&gt;
&lt;li&gt;Skills come from a dictionary match first. An optional AI pass can refine them later, along with a short summary and a flag for postings that look like scams.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The same job, five times
&lt;/h2&gt;

&lt;p&gt;Popular roles are posted on several boards, and nobody wants to scroll past the same opening five times. The importer deals with that in three steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Same source, same ID&lt;/strong&gt; as a job already saved: update it in place, so its page address never changes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Different source, same job&lt;/strong&gt;: skip it. The check is a SHA-1 hash of the title, company and location, lowercased and with spaces collapsed, so "Senior  Engineer" and "senior engineer" match.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A new job&lt;/strong&gt; gets a readable address made from its title, company and place, with -2 or -3 added only when two jobs would collide.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Every job is a real page
&lt;/h2&gt;

&lt;p&gt;Each job has its own page under &lt;code&gt;/jobs/&lt;/code&gt;, with JobPosting structured data so Google can show it in job search. A posting counts as valid for 30 days from the day it was posted. The dashboard, the API and the sign-in pages are kept out of search in &lt;code&gt;robots.txt&lt;/code&gt;, so only the public pages get indexed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Running the import on Vercel
&lt;/h2&gt;

&lt;p&gt;The import starts from a scheduled call to an endpoint that needs a secret. When Redis is available, the endpoint puts the run on a BullMQ queue and returns. When it isn't, it answers at once and keeps working in the background with Vercel's &lt;code&gt;waitUntil&lt;/code&gt;. That second path exists because some external cron services stop waiting after about 30 seconds, and an import can take longer than that. Without it, a run that worked would still be logged as a failure.&lt;/p&gt;

&lt;h2&gt;
  
  
  One AI client, three providers
&lt;/h2&gt;

&lt;p&gt;Every AI feature goes through one small client. Groq, Google Gemini and OpenAI all accept the same chat-completions format, so a single code path serves all three. The client tries the providers that have a key, in order (Groq with Llama 3.3 70B first, then Gemini, then OpenAI), retries once on a rate limit or a server error, and then moves on to the next.&lt;/p&gt;

&lt;p&gt;The rule I care about most is at the end of that client: if every provider fails, it raises a clear error. It never returns sample data. A made-up résumé score is worse than no score, because people act on it.&lt;/p&gt;

&lt;p&gt;Each feature asks the model for one JSON object with fixed keys. The résumé analysis, for example, returns a score from 0 to 100, an ATS score, the skills it found, strengths, specific improvements and a short summary. Models sometimes wrap JSON in code fences or add a sentence before it, so a small function strips that and takes the outermost object before parsing. If it still doesn't parse, the error says so.&lt;/p&gt;

&lt;h2&gt;
  
  
  Reading a résumé
&lt;/h2&gt;

&lt;p&gt;Résumés arrive as PDFs. The server pulls the text out with &lt;code&gt;pdf-parse&lt;/code&gt; and keeps the file in a private storage bucket. A scanned résumé is only a picture of text, and the text models can't read pictures, so NexeraAI says that plainly and asks for a text PDF instead of guessing.&lt;/p&gt;

&lt;h2&gt;
  
  
  An assistant that only suggests real roles
&lt;/h2&gt;

&lt;p&gt;The Nexera AI chat can talk about anything, but it is built for careers. With each question it gets today's date, recent web results when they help, and the open roles on Nexera that fit the question. Its instructions say to recommend only from that list, by exact title, and never to invent a role. The buttons under an answer are plain search queries, so a tap takes you to real listings.&lt;/p&gt;

&lt;p&gt;Job matching works the same way. The model compares your résumé with one job's description and returns a match percentage, the skills you have, the skills you are missing, a fit level and a short explanation. I asked for the missing skills and the explanation on purpose: a percentage alone doesn't tell you what to do next.&lt;/p&gt;

&lt;h2&gt;
  
  
  The stack
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Next.js (App Router) and TypeScript, Tailwind CSS v4 with shadcn/ui, and Framer Motion and GSAP for motion&lt;/li&gt;
&lt;li&gt;PostgreSQL through Prisma, with résumés in a private Supabase storage bucket&lt;/li&gt;
&lt;li&gt;Auth.js for sign-in with Google, GitHub, or email and a password&lt;/li&gt;
&lt;li&gt;BullMQ and Redis for the import queue, when Redis is there&lt;/li&gt;
&lt;li&gt;Razorpay for the paid plans&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What I would tell someone building one
&lt;/h2&gt;

&lt;p&gt;Start with the data, not the prompts. Clean listings, no duplicates and page addresses that never change made every AI feature after them simpler, and they are also what search engines see. And decide early what the product does when the AI is down. For NexeraAI the answer is to say so.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This post first appeared on my blog at &lt;a href="https://shivamjha.dev/blog/how-i-built-nexeraai.html" rel="noopener noreferrer"&gt;shivamjha.dev&lt;/a&gt;. I'm Shivam Anil Jha, an AI architect and full-stack developer from Mumbai. I also build &lt;a href="https://clonix-tau.vercel.app/" rel="noopener noreferrer"&gt;Clonix&lt;/a&gt; and &lt;a href="https://github.com/shivamj56/orbit-os" rel="noopener noreferrer"&gt;Orbit OS&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>nextjs</category>
      <category>ai</category>
      <category>typescript</category>
      <category>showdev</category>
    </item>
    <item>
      <title>I built a phone app that debloats Google TV over Wi-Fi, with an ADB client in pure Dart</title>
      <dc:creator>Shivam Anil Jha</dc:creator>
      <pubDate>Mon, 16 Nov 2026 07:19:49 +0000</pubDate>
      <link>https://dev.to/shivamj56/i-built-a-phone-app-that-debloats-google-tv-over-wi-fi-with-an-adb-client-in-pure-dart-35lk</link>
      <guid>https://dev.to/shivamj56/i-built-a-phone-app-that-debloats-google-tv-over-wi-fi-with-an-adb-client-in-pure-dart-35lk</guid>
      <description>&lt;p&gt;Our TV's home screen was half ads. A big sponsored banner, rows nobody asked for, and dozens of preinstalled apps waking up with the TV. Every fix I found started with "connect a PC and run these ADB commands", with no record of what changed and no easy way back.&lt;/p&gt;

&lt;p&gt;So I built &lt;strong&gt;Orbit OS&lt;/strong&gt;: a free, open-source (GPL-3.0) Android app that sets up a Google TV or Android TV from your phone over the home Wi-Fi. No PC, no cables, no commands, and one tap of Undo puts the TV back exactly as it came.&lt;/p&gt;

&lt;p&gt;This post is about how it works under the hood.&lt;/p&gt;

&lt;h2&gt;
  
  
  What "debloating" a TV actually means
&lt;/h2&gt;

&lt;p&gt;There's no magic. Android already has the tool:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb shell pm disable-user --user 0 com.google.android.feedback   &lt;span class="c"&gt;# switch an app off&lt;/span&gt;
adb shell pm &lt;span class="nb"&gt;enable&lt;/span&gt; com.google.android.feedback                  &lt;span class="c"&gt;# and back on&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A disabled app stays on the system partition and never runs. &lt;code&gt;pm uninstall --user 0&lt;/code&gt; also works, but it is harder to undo and gains nothing. Orbit only ever disables.&lt;/p&gt;

&lt;p&gt;The hard part isn't the command. It's knowing &lt;strong&gt;what is safe to switch off on this TV&lt;/strong&gt;, doing it &lt;strong&gt;without a computer&lt;/strong&gt;, and being able to &lt;strong&gt;undo everything&lt;/strong&gt; later.&lt;/p&gt;

&lt;h2&gt;
  
  
  An ADB client in pure Dart
&lt;/h2&gt;

&lt;p&gt;The phone app is Flutter. Instead of shipping a native &lt;code&gt;adb&lt;/code&gt; binary, it speaks the ADB wire protocol itself, in one ~540-line Dart file. ADB over TCP is a small message protocol on port 5555. Every message has a 24-byte header (command, two arguments, payload length, checksum, and the command XOR 0xffffffff), followed by the payload:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Command&lt;/th&gt;
&lt;th&gt;Meaning&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;CNXN&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;connect, exchange version and max payload size&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;AUTH&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;the TV sends a token; the phone signs it, or sends its public key&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;OPEN&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;open a stream, for example &lt;code&gt;shell:pm list packages&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;WRTE&lt;/code&gt; / &lt;code&gt;OKAY&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;data, and flow control&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;CLSE&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;close the stream&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The first time, the TV doesn't know the phone's key, so the phone sends its RSA public key and the TV shows the familiar &lt;strong&gt;"Allow USB debugging?"&lt;/strong&gt; dialog. The user presses Allow once with the TV remote. Generating the key pair and signing the TV's token is a small Kotlin helper; everything else is Dart.&lt;/p&gt;

&lt;p&gt;One limitation: some newer devices only offer &lt;em&gt;Wireless debugging&lt;/em&gt; with pairing codes, which is TLS-based. Orbit doesn't support those yet.&lt;/p&gt;

&lt;h2&gt;
  
  
  The scripts that change the TV
&lt;/h2&gt;

&lt;p&gt;The phone never fires commands blindly. It pushes two shell scripts to the TV, &lt;code&gt;orbit-setup.sh&lt;/code&gt; and &lt;code&gt;orbit-undo.sh&lt;/code&gt;, and drives them. Lines starting with &lt;code&gt;@orbit&lt;/code&gt; are a tiny protocol for the phone (&lt;code&gt;step&lt;/code&gt;, &lt;code&gt;progress&lt;/code&gt;, &lt;code&gt;undo&lt;/code&gt;, &lt;code&gt;done&lt;/code&gt;); everything else is log. That's how the app shows every step live.&lt;/p&gt;

&lt;p&gt;Safety comes from the &lt;strong&gt;order of operations&lt;/strong&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Write the undo record first.&lt;/strong&gt; Before any change, the script records it in &lt;code&gt;undo.txt&lt;/code&gt; and streams it to the phone, together with the TV's model, Android version and app list.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Never leave the TV without a home screen.&lt;/strong&gt; The stock home is switched off only &lt;em&gt;after&lt;/em&gt; Orbit Home, our launcher, has started. Otherwise a TV can boot to a blank screen.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One list decides everything.&lt;/strong&gt; A catalogue in the code groups packages into &lt;em&gt;ads and tracking&lt;/em&gt;, &lt;em&gt;apps nobody uses&lt;/em&gt;, &lt;em&gt;woken at every start&lt;/em&gt; and &lt;em&gt;the stock home screen&lt;/em&gt;, plus a &lt;code&gt;neverOff&lt;/code&gt; list: Play services, the Play Store, the TV Remote Service, the maker's input and remote services. On TVs it hasn't been tested on, Orbit offers only the changes it knows are safe anywhere.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Before anything changes, the user answers five plain questions (do you cast? use voice search? play from USB?) and sees the whole plan.&lt;/p&gt;

&lt;h2&gt;
  
  
  Things I learned the hard way
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;On Xiaomi TVs, &lt;code&gt;com.mitv.overseaservice&lt;/code&gt; looks like bloat but keeps the remote paired after standby. It's on the never-touch list now.&lt;/li&gt;
&lt;li&gt;On the tested TV, Google TV's setup app has to be switched off &lt;strong&gt;together&lt;/strong&gt; with the stock home, or it brings the stock home back.&lt;/li&gt;
&lt;li&gt;Google TV's built-in &lt;em&gt;Apps only mode&lt;/em&gt; hides recommendations, but Google says sponsored content can still appear. The only real fix for the banner is a different home screen.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Numbers from the test TV
&lt;/h2&gt;

&lt;p&gt;On a Xiaomi TV A Pro with Google TV 14: 32 preinstalled apps switched off (none deleted), Orbit Home's Control Center at 60 fps with 0 janky frames in 400, and a full setup in about five minutes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it, or read the code
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Google Play: &lt;a href="https://play.google.com/store/apps/details?id=dev.shiv.orbit" rel="noopener noreferrer"&gt;https://play.google.com/store/apps/details?id=dev.shiv.orbit&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Code (GPL-3.0): &lt;a href="https://github.com/shivamj56/orbit-os" rel="noopener noreferrer"&gt;https://github.com/shivamj56/orbit-os&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Prefer doing it by hand? &lt;a href="https://orbit-os-tv.vercel.app/guides/debloat-google-tv.html" rel="noopener noreferrer"&gt;https://orbit-os-tv.vercel.app/guides/debloat-google-tv.html&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It works on Google TV and Android TV sets (not Samsung, LG, Roku or Fire TV). If you have a different TV, I'd love a report, and contributions are welcome. What would you switch off on yours?&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>android</category>
      <category>opensource</category>
      <category>flutter</category>
    </item>
  </channel>
</rss>
