<?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: Nikhil Kumar</title>
    <description>The latest articles on DEV Community by Nikhil Kumar (@nikhilkumarswe).</description>
    <link>https://dev.to/nikhilkumarswe</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%2F4199624%2F44cd4b34-925a-41b0-aad8-af21ff8498d4.png</url>
      <title>DEV Community: Nikhil Kumar</title>
      <link>https://dev.to/nikhilkumarswe</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/nikhilkumarswe"/>
    <language>en</language>
    <item>
      <title>Virtual Ofrenda &amp; Story Alchemist: remembering someone with an AI that never leaves your device</title>
      <dc:creator>Nikhil Kumar</dc:creator>
      <pubDate>Sun, 01 Nov 2026 06:48:22 +0000</pubDate>
      <link>https://dev.to/nikhilkumarswe/virtual-ofrenda-story-alchemist-remembering-someone-with-an-ai-that-never-leaves-your-device-5254</link>
      <guid>https://dev.to/nikhilkumarswe/virtual-ofrenda-story-alchemist-remembering-someone-with-an-ai-that-never-leaves-your-device-5254</guid>
      <description>&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;When someone dies, the memories that matter most are rarely the grand ones. They are the yellow cardigan, the Sunday porch, the way someone laughed in a thunderstorm. I wanted to build a small, quiet place to put those down.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Virtual Ofrenda &amp;amp; Story Alchemist&lt;/strong&gt; is a private web app. You type or speak a few memories of someone you miss. A small open-weight model, running on your own device, turns them into a short tribute story and a poem, and builds a candlelit digital &lt;em&gt;ofrenda&lt;/em&gt; (altar) from the things you mentioned: candles, marigolds, a favourite book, a hot mug of cider.&lt;/p&gt;

&lt;p&gt;The theme asked for something that honors the past in the spirit of Día de los Muertos. The ofrenda tradition is much bigger than any app, so I borrowed only its simplest idea: that people stay with us while we tell their stories, and that we make small, physical places to hold them.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu2a1pqsq6642xr81gf0u.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu2a1pqsq6642xr81gf0u.jpg" alt=" " width="800" height="363"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It has three parts:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Voice-to-Memory.&lt;/strong&gt; Record a memory in the browser and it is transcribed locally with whisper.cpp. You can edit the text, or just type.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Story Alchemist.&lt;/strong&gt; One tribute story and one poem, made from what you said.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Altar Decorator.&lt;/strong&gt; The objects in your memory land on a three-tier altar and fade in, one by one.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The page says "Runs on your device" because it does. There are no accounts, no cloud APIs and no tracking. I tested it in airplane mode.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/wHRRkFlyoqc" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;The video shows the typing path from start to finish: a memory about a fictional woman named Eleanor, then the altar, the story and the poem. I recorded it on the same phone the app runs on, and the waiting in the video is shortened, because generating a tribute takes a few minutes on this hardware.&lt;/p&gt;

&lt;p&gt;I left voice input out of the recording. It works, but on my phone Whisper needed about two minutes for an 11-second clip, so typing is the dependable path there.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F73w0mel9n48r7eyuvte9.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F73w0mel9n48r7eyuvte9.png" alt=" " width="800" height="382"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/nikhil-kumar-swe" rel="noopener noreferrer"&gt;
        nikhil-kumar-swe
      &lt;/a&gt; / &lt;a href="https://github.com/nikhil-kumar-swe/virtual-ofrenda" rel="noopener noreferrer"&gt;
        virtual-ofrenda
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Virtual Ofrenda &amp;amp; Story Alchemist&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;A private web app for remembering someone you have lost. You speak or type a few memories of them, and a small open-weight AI running &lt;strong&gt;on your own device&lt;/strong&gt; turns those memories into a tribute story and a short poem, then decorates a candlelit digital &lt;em&gt;ofrenda&lt;/em&gt; (altar) with the things you mentioned: candles, marigolds, a favourite book, a bright yellow raincoat.&lt;/p&gt;
&lt;p&gt;Nothing leaves the device. There are no cloud APIs, no accounts and no tracking. The whole project was built on a low-end device with 2 GB of RAM.&lt;/p&gt;
&lt;p&gt;Built for the &lt;strong&gt;DEV Hacktoberfest 2026, Week 4 challenge ("Ghost in the Machine")&lt;/strong&gt;.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Demo video:&lt;/strong&gt; &lt;a href="https://www.youtube.com/watch?v=wHRRkFlyoqc" rel="nofollow noopener noreferrer"&gt;https://www.youtube.com/watch?v=wHRRkFlyoqc&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What it does&lt;/h2&gt;
&lt;/div&gt;
&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Voice-to-Memory.&lt;/strong&gt; Record a memory in the browser. It is transcribed locally with whisper.cpp (&lt;code&gt;tiny.en&lt;/code&gt;) and lands in a text box where you can edit it. Typing works just as well and is the main…&lt;/li&gt;
&lt;/ol&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/nikhil-kumar-swe/virtual-ofrenda" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The repository is MIT licensed and has setup steps, the model list and the commands to start each server. The model files are not included, so the README tells you how to download them.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;The constraint.&lt;/strong&gt; I built this on an Android phone with 2 GB of RAM and a 32-bit ARM processor, using Termux. My first plan was a 7 to 8 billion parameter model, and then a tiny Gemma through Ollama. Ollama does not run on 32-bit ARM, and a 7B model would not have fit in that memory anyway.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The open pieces.&lt;/strong&gt; So I went smaller and more direct:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Gemma 3 270M (instruction-tuned, Q8_0, about 300 MB)&lt;/strong&gt; rewrites the story, writes the poem and suggests the altar items. It runs through &lt;strong&gt;llama.cpp&lt;/strong&gt;'s &lt;code&gt;llama-server&lt;/code&gt;, built from source on the phone with &lt;code&gt;-DGGML_NATIVE=OFF -DGGML_LLAMAFILE=OFF&lt;/code&gt;. It generates about 6.7 tokens per second.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Whisper tiny.en&lt;/strong&gt; through &lt;strong&gt;whisper.cpp&lt;/strong&gt;'s &lt;code&gt;whisper-server&lt;/code&gt; turns speech into text.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Laravel&lt;/strong&gt; with Blade, Tailwind and SQLite runs the app. A queued job calls the model, and the browser polls for the result, so the story, the poem and the altar items appear as they arrive.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Take the model away and the app has no story, no poem and no altar choices to show. Everything else is plumbing around it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg15u4my10pfviefp8koi.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg15u4my10pfviefp8koi.jpg" alt=" " width="800" height="299"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Making a tiny model behave.&lt;/strong&gt; A 270M model will happily make things up, and in a tribute that matters. So the app puts guard rails around it. They are uneven, and I want to be exact about where they are:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The story is checked sentence by sentence.&lt;/strong&gt; The app takes the first six sentences of your memory and asks the model to tidy each one. A rewritten sentence is kept only if it uses words from the original sentence and covers most of it. Otherwise the app falls back to your own sentence. This is why the stories never invented a fact in my tests, and also why they mostly read like your own words.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The altar is checked against your text.&lt;/strong&gt; The model suggests objects, a keyword list adds a few more, and each item must appear in your memory as a whole word. Candles and marigolds are always added. Near-duplicate items are merged, so "green shards" and "frosted green shards" become one.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The poem is not checked.&lt;/strong&gt; The model writes four short lines, the app tidies them and adds a title line and a closing line. Nothing verifies that the lines come from your memory, so this is where invented imagery can slip in.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Failures are friendly.&lt;/strong&gt; If a model server is offline, you get a clear message and a retry button, not a crash.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;What still goes wrong.&lt;/strong&gt; I tested it on three memories of different lengths (about 38, 85 and 130 words). These are the honest results, plus a few things I found later when I read my own code:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The stories added no invented facts, but mostly because they repeat your own words.&lt;/strong&gt; On the Eleanor memory, the story was almost word for word what I had typed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The narrator can change.&lt;/strong&gt; My story prompt asks the model to write in the first person, and my check does not notice a change of voice. So a short memory about a man named Marcus came back as "I always carried a pocketful of sea glass" instead of "Marcus always carried…".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The poem can add imagery.&lt;/strong&gt; It invented no people or events in my tests, but it added phrases like "secrets held tight" that were never in the memory.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Altar items are sometimes ideas, not objects.&lt;/strong&gt; The model picked "quiet invitation" for Eleanor's altar. It passed my check because those words do appear in her memory. It shows up with the default dove icon, in the screenshot above.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Long memories are cut.&lt;/strong&gt; Only the first 1200 characters, and the first six sentences for the story, are used, and the app does not say so.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A bug I found after the deadline.&lt;/strong&gt; If the poem step gets too little from the model, the app falls back to four fixed lines that I wrote while testing one particular memory. For anyone else those lines would be wrong. It is first on my fix list.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Here is Eleanor's poem, as the model wrote it:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;Yellow cardigan scent&lt;/em&gt;&lt;br&gt;
&lt;em&gt;Lavender whispers on porch&lt;/em&gt;&lt;br&gt;
&lt;em&gt;Bird wings, watercolor splashes&lt;/em&gt;&lt;br&gt;
&lt;em&gt;Ending's a quiet invitation&lt;/em&gt;&lt;br&gt;
&lt;em&gt;Remembered with love.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;It is not Rilke. This time every image in it came from the memory. Nothing guarantees that, which is exactly why I am telling you about the other cases.&lt;/p&gt;

&lt;p&gt;I built the project between Oct 29 and the morning of Nov 1 (IST), in 11 commits, all before the contest deadline.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;A memory of someone you have lost is some of the most private writing a person will ever do. If this app had to call a hosted model, every one of those memories would leave the device and sit on someone else's server. I did not want to build that, and I would not want to use it.&lt;/p&gt;

&lt;p&gt;Open weights made a different design possible. Gemma is small enough to run locally, and open enough that I could serve it myself, with no key, no account and no network. When the internet is off, it still works.&lt;/p&gt;

&lt;p&gt;Open tools did the rest of the work. llama.cpp and whisper.cpp run on hardware that the friendlier tools do not support. My phone's 32-bit chip could not run Ollama, but it could run llama.cpp built from source. That is the difference between "this idea needs a data center" and "this idea fits in a pocket".&lt;/p&gt;

&lt;p&gt;Openness also makes the project easy to improve. If you have a laptop, you can serve a larger open model with llama.cpp and point the same app at it by changing the model name and URL in &lt;code&gt;.env&lt;/code&gt;. The app itself does not need to change. Nobody has to ask permission.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Spirit of It
&lt;/h2&gt;

&lt;p&gt;The constraint I started with turned out to be the idea. A small model cannot be trusted to be clever, so I made the app's promise modest: &lt;em&gt;using only what you tell us&lt;/em&gt;. It does not try to write a beautiful eulogy for someone it never met. It gives your own words back, arranged with care, and puts a few candles around them.&lt;/p&gt;

&lt;p&gt;I should be honest about how well that promise holds. It holds best for the story and the altar, which are checked against your words. It holds least for the poem, where a tiny model still wanders. Seeing that clearly, and saying it, felt more in keeping with an ofrenda than polishing it away.&lt;/p&gt;

&lt;p&gt;There is plenty I would still like to do: fix the fallback poem, check the poem against the memory, write the story in the third person, add a page that lists past tributes (today you return to one by its link), prompt starters for people who do not know where to begin, and a faster speech model. But the core is there, and it fits on a phone with 2 GB of RAM.&lt;/p&gt;

&lt;p&gt;If you try it and it gets something wrong, I would like to hear about it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Gemma.&lt;/strong&gt; The project is built around Google's open-weight Gemma 3 270M model, which does the writing and choosing in every tribute: the story, the poem and the altar items. It runs locally through llama.cpp, with no hosted API anywhere in the app.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Edited after the contest deadline: I read the code again and corrected how I had described the poem and the story checks, and added the limits and the fallback-poem bug above. The submitted code was not changed.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>I built a lazy job-application helper with Gemma</title>
      <dc:creator>Nikhil Kumar</dc:creator>
      <pubDate>Mon, 26 Oct 2026 06:54:26 +0000</pubDate>
      <link>https://dev.to/nikhilkumarswe/i-built-a-lazy-job-application-helper-with-gemma-2lib</link>
      <guid>https://dev.to/nikhilkumarswe/i-built-a-lazy-job-application-helper-with-gemma-2lib</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;The task I dread most when applying for jobs is rewriting my cover letter and resume for every single job description. So I built the &lt;strong&gt;Lazy Job-Application Helper&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;You paste your resume as plain text and paste a job description. Gemma, an open-weight model from Google, sends back four things:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a tailored cover letter&lt;/li&gt;
&lt;li&gt;the keywords the job asks for that your resume doesn't show&lt;/li&gt;
&lt;li&gt;3 to 5 rewritten resume bullets&lt;/li&gt;
&lt;li&gt;5 likely interview questions&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Everything is saved to a history page where you can reopen or delete it. Each account can only see its own applications.&lt;/p&gt;

&lt;p&gt;It gives suggestions only. It doesn't write a whole resume, and it doesn't apply for you. The rule I cared about most is that &lt;strong&gt;the model may only use facts that are in your resume.&lt;/strong&gt; It must never invent skills, tools or numbers, because a cover letter that claims experience you don't have is worse than no cover letter.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;Video: &lt;a href="https://www.youtube.com/watch?v=cR_yfAr1JYI" rel="noopener noreferrer"&gt;https://www.youtube.com/watch?v=cR_yfAr1JYI&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;A few honest notes about the demo:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;I recorded it on my phone, and it runs locally (that's the &lt;code&gt;127.0.0.1&lt;/code&gt; in the address bar).&lt;/li&gt;
&lt;li&gt;I paused the recording while Gemma was generating, so the video skips the wait. On the free tier a real run takes up to about 2 minutes.&lt;/li&gt;
&lt;li&gt;I did not deploy a public link on purpose. A public demo would let strangers spend my free API quota and send their resumes to a third party.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I tested it with mismatched sample resumes on purpose. When I gave it a DevOps resume and a Frontend (ReactJS + InertiaJS) job, it listed ReactJS, InertiaJS, Tailwind CSS and Vite as missing keywords. Its interview questions asked how a DevOps background prepares me for a frontend role. It did not pretend I had React experience, and that is exactly what I wanted.&lt;/p&gt;

&lt;p&gt;The screenshot below shows the same idea the other way round: a front-end resume against a Java backend job. Java, Spring and Unix come back as missing keywords, and the rewritten bullets keep the real numbers from the resume.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F96tveul1hup81lzs3dh9.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F96tveul1hup81lzs3dh9.png" alt=" " width="799" height="389"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1gydstrtt3jrikdv8wct.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1gydstrtt3jrikdv8wct.png" alt=" " width="799" height="413"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/nikhil-kumar-swe" rel="noopener noreferrer"&gt;
        nikhil-kumar-swe
      &lt;/a&gt; / &lt;a href="https://github.com/nikhil-kumar-swe/lazy-job-helper" rel="noopener noreferrer"&gt;
        lazy-job-helper
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Lazy Job-Application Helper&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;Paste your resume and a job description. Gemma gives you a tailored cover letter, the keywords your resume is missing, a few resume bullet rewrites, and five likely interview questions. Everything is saved to a private history page.&lt;/p&gt;
&lt;p&gt;Built for the &lt;strong&gt;DEV Hacktoberfest Open-Source AI Challenge: Week 3 ("Slow and Steady")&lt;/strong&gt;. The theme is open-weight AI that does a dreaded task for you, and writing job applications is mine.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Repo: &lt;a href="https://github.com/nikhil-kumar-swe/lazy-job-helper" rel="noopener noreferrer"&gt;https://github.com/nikhil-kumar-swe/lazy-job-helper&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What it does&lt;/h2&gt;
&lt;/div&gt;
&lt;ol&gt;
&lt;li&gt;Sign up and log in (Laravel Breeze).&lt;/li&gt;
&lt;li&gt;Paste your resume as plain text and paste the job description.&lt;/li&gt;
&lt;li&gt;A hosted Gemma model returns
&lt;ul&gt;
&lt;li&gt;a tailored cover letter&lt;/li&gt;
&lt;li&gt;missing keywords (things the job asks for that your resume does not show)&lt;/li&gt;
&lt;li&gt;3 to 5 rewritten resume bullets&lt;/li&gt;
&lt;li&gt;5 interview questions&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;Results are saved. You can view or delete them from the History page.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;&lt;strong&gt;Suggestions only.&lt;/strong&gt; It does not generate a full resume, and…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/nikhil-kumar-swe/lazy-job-helper" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;This was my first time calling an AI model from code.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;App:&lt;/strong&gt; Laravel with Breeze (login), Blade and Tailwind CSS, with SQLite for development.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; &lt;code&gt;gemma-4-31b-it&lt;/code&gt;, served through the Gemini API on Google AI Studio's free tier.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One class for all model calls:&lt;/strong&gt; &lt;code&gt;ModelClient.php&lt;/code&gt; reads the API key, URL and model name from &lt;code&gt;.env&lt;/code&gt;. It asks for a JSON answer, pulls out the JSON, checks that all four fields are present, and retries once if the output is unreadable. It also retries on temporary API errors and shows a friendly message if the call fails.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Safety basics:&lt;/strong&gt; keys stay out of git, submissions are rate limited, and other users' result pages return a 404.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Things that went wrong along the way:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The first model name I used returned a 404. I asked the API which models my key could use and found the right one.&lt;/li&gt;
&lt;li&gt;Gemma 4 returns its "thinking" as a separate part of the reply. My first version would have shown that as the answer, so I filter it out.&lt;/li&gt;
&lt;li&gt;Google's API returned a 500 error once, which is why the retries exist.&lt;/li&gt;
&lt;li&gt;The Vite build failed on Termux until I added esbuild.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The prompt took four rounds, and I checked every round against the resume text by hand:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;First version:&lt;/strong&gt; no invented skills, but it added filler like "successfully" and "to enhance team capabilities".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ban additions:&lt;/strong&gt; no more filler, but the bullets were nearly copies of my resume.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep numbers exact:&lt;/strong&gt; the numbers survived, but one bullet still came out identical.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Final version:&lt;/strong&gt; pick the bullets most relevant to the job, start with a strong verb, put the result first, keep the same meaning, never copy word for word, and add no new facts.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;With the final prompt, the rewritten bullets matched facts in the resume, including the numbers (+200, +24%, $5,000, 5,500 lines of code and a team of 4). The cost of being strict is that the rewrites stay fairly close to the original wording. I chose honest over flashy.&lt;/p&gt;

&lt;p&gt;The cover letter and interview questions are less tightly controlled than the bullets. In some runs they still add soft phrases like "proven track record" or "extensive experience" that my resume doesn't state. That is why the result page tells you to check every line before using it.&lt;/p&gt;

&lt;p&gt;I also used an AI assistant to help plan the build and debug errors, and I ran and tested every step myself.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;I can't run a model this big myself.&lt;/strong&gt; A 31B-parameter model is too large for the hardware I have. A hosted open-weight model let me use Gemma anyway, at no cost to me.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I'm not locked in to one model.&lt;/strong&gt; All model calls go through one class and three &lt;code&gt;.env&lt;/code&gt; values. A different Gemma size is a &lt;code&gt;.env&lt;/code&gt; change, and a different provider means editing one file.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Open weights make a private option possible.&lt;/strong&gt; Someone with a stronger machine could run Gemma locally, for example with Ollama, and keep their resume on their own computer. I did not build or test that, so I'm not claiming it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The honest part: my version uses Google's hosted free tier. Resume text goes to Google, and on the free tier Google may use it to improve its models. That is why the form warns you not to paste phone numbers, addresses or ID numbers, and why I don't call this version private. For a tool that handles resumes, being able to move to a local model is the real reason open weights matter to me.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Lazy Payoff
&lt;/h2&gt;

&lt;p&gt;I timed myself writing a cover letter and new resume bullets for one job by hand. It took &lt;strong&gt;10 minutes&lt;/strong&gt;. The app returned its answer in about 2 minutes of waiting, and that answer also included the missing keywords and 5 interview questions I didn't write by hand.&lt;/p&gt;

&lt;p&gt;This was one task, one person and one sample resume, and I didn't stopwatch the app, so treat it as a rough comparison. Nobody else has used it yet, and I still need to read and check everything it writes before using any of it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Best Use of Gemma&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
  </channel>
</rss>
