<?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: Pushpraj Patel</title>
    <description>The latest articles on DEV Community by Pushpraj Patel (@vilaspatel938).</description>
    <link>https://dev.to/vilaspatel938</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%2F4155987%2F19bc20f4-d1b6-473f-97aa-7f66512d682f.png</url>
      <title>DEV Community: Pushpraj Patel</title>
      <link>https://dev.to/vilaspatel938</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/vilaspatel938"/>
    <language>en</language>
    <item>
      <title>🕯️ ECHOES — Memories Never Die | A Local AI Memory Keeper</title>
      <dc:creator>Pushpraj Patel</dc:creator>
      <pubDate>Sat, 31 Oct 2026 20:24:44 +0000</pubDate>
      <link>https://dev.to/vilaspatel938/echoes-memories-never-die-a-local-ai-memory-keeper-55d7</link>
      <guid>https://dev.to/vilaspatel938/echoes-memories-never-die-a-local-ai-memory-keeper-55d7</guid>
      <description>&lt;p&gt;Some memories deserve more than a place in our photo gallery. They deserve to be remembered, revisited, and told again.&lt;/p&gt;

&lt;p&gt;For the Hacktoberfest Open-Source AI Challenge: Week 4 — Ghost in the Machine, I built ECHOES, a local-first AI memory keeper powered by Gemma 4.&lt;/p&gt;

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

&lt;p&gt;ECHOES is a personal AI memory assistant that allows people to preserve meaningful memories, stories, and photographs, then ask questions about them using a locally running AI model.&lt;/p&gt;

&lt;p&gt;The idea came from the thought that our memories are scattered across old photographs, notes, and conversations. What if we could bring them together and have an AI help us revisit those moments?&lt;/p&gt;

&lt;p&gt;With ECHOES, users can:&lt;/p&gt;

&lt;p&gt;🕯️ Preserve memories with a person's name, title, story, and photograph.&lt;/p&gt;

&lt;p&gt;📖 Browse and search saved memories in a Memory Vault.&lt;/p&gt;

&lt;p&gt;💬 Ask questions about their saved memories.&lt;/p&gt;

&lt;p&gt;🧠 Get responses generated by Gemma 4.&lt;/p&gt;

&lt;p&gt;🔒 Keep their memories and AI interactions local.&lt;/p&gt;

&lt;p&gt;I interpreted Ghost in the Machine as giving memories a voice through AI — not by pretending to bring someone back, but by helping people reconnect with the stories they have preserved.&lt;/p&gt;

&lt;p&gt;Code&lt;/p&gt;

&lt;p&gt;GitHub Repository: &lt;a href="https://github.com/vikasfine17/echoes" rel="noopener noreferrer"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The project is built with Python and Streamlit, with SQLite for local storage.&lt;/p&gt;

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

&lt;p&gt;I wanted to keep the architecture simple and make local AI the core of the application.&lt;/p&gt;

&lt;p&gt;Tech stack:&lt;/p&gt;

&lt;p&gt;Python — Application logic.&lt;/p&gt;

&lt;p&gt;Streamlit — User interface.&lt;/p&gt;

&lt;p&gt;SQLite — Stores memories and photographs locally.&lt;/p&gt;

&lt;p&gt;Ollama — Runs the AI model locally.&lt;/p&gt;

&lt;p&gt;Gemma 4 (gemma4) — Powers the conversational memory assistant.&lt;/p&gt;

&lt;p&gt;When someone asks ECHOES a question, the application retrieves the saved memories from SQLite and passes their content to Gemma through Ollama's local API. Gemma then generates a response based on that information.&lt;/p&gt;

&lt;p&gt;I also instructed the model not to invent memories or details when the answer isn't available in the saved information.&lt;/p&gt;

&lt;p&gt;The entire AI workflow runs on the user's own machine without requiring a hosted AI API.&lt;/p&gt;

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

&lt;p&gt;For a project dealing with personal memories, privacy isn't just another feature. It's fundamental.&lt;/p&gt;

&lt;p&gt;I didn't want someone's personal stories and photographs to depend on a third-party cloud AI service.&lt;/p&gt;

&lt;p&gt;Using open-weight Gemma 4 and Ollama made it possible to:&lt;/p&gt;

&lt;p&gt;Keep data local: Memories remain in the user's local SQLite database.&lt;/p&gt;

&lt;p&gt;Improve privacy: Questions and memory context are sent to a locally running model rather than a hosted AI API.&lt;/p&gt;

&lt;p&gt;Reduce dependency: The application doesn't depend on a paid AI API subscription.&lt;/p&gt;

&lt;p&gt;Experiment freely: Open-weight models make it possible to explore different models and customize the experience.&lt;/p&gt;

&lt;p&gt;Build with accessibility in mind: Once the model is downloaded, the application can run without an internet connection, provided the local services are running.&lt;/p&gt;

&lt;p&gt;Open innovation gave me the freedom to build an AI experience around personal data without making a cloud provider the center of the project.&lt;/p&gt;

&lt;p&gt;My Agent Session&lt;/p&gt;

&lt;p&gt;I haven't included a DevRelay session for this project.&lt;/p&gt;

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

&lt;p&gt;We take hundreds of photographs, save thousands of messages, and collect countless little moments.&lt;/p&gt;

&lt;p&gt;But sometimes, we forget the stories behind them.&lt;/p&gt;

&lt;p&gt;ECHOES is my small attempt to make those stories easier to revisit.&lt;/p&gt;

&lt;p&gt;Imagine asking:&lt;/p&gt;

&lt;p&gt;"What did my grandfather teach me when I was young?"&lt;/p&gt;

&lt;p&gt;And getting an answer from a story you personally preserved.&lt;/p&gt;

&lt;p&gt;It's not about recreating people or pretending AI can replace them. It's about keeping the memories they left behind accessible.&lt;/p&gt;

&lt;p&gt;People may leave, but their stories don't have to disappear.&lt;/p&gt;

&lt;p&gt;Prize Categories&lt;/p&gt;

&lt;p&gt;Best Use of Gemma&lt;/p&gt;

&lt;p&gt;Built with curiosity, Python, and open-weight AI.&lt;/p&gt;

&lt;h1&gt;
  
  
  devchallenge #hf26challenge #hacktoberfest #AI #Gemma #Ollama #OpenSourceAI #Python
&lt;/h1&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Sloth Sorter</title>
      <dc:creator>Pushpraj Patel</dc:creator>
      <pubDate>Thu, 22 Oct 2026 18:59:59 +0000</pubDate>
      <link>https://dev.to/vilaspatel938/sloth-sorter-29oi</link>
      <guid>https://dev.to/vilaspatel938/sloth-sorter-29oi</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;&lt;strong&gt;Sloth Sorter&lt;/strong&gt; is a lazy file tidier. Point it at a messy folder (my Downloads folder was the victim) and it reads each file, decides where it belongs, gives it a clean descriptive name, and files it away. You review a plan and approve it with one flag.&lt;/p&gt;

&lt;p&gt;It saves the chore everyone puts off: dragging &lt;code&gt;Screenshot (47).png&lt;/code&gt;, &lt;code&gt;invoice_final_v2.pdf&lt;/code&gt; and &lt;code&gt;download(3).zip&lt;/code&gt; into folders and renaming them. Anyone with a cluttered Downloads folder gets to be a little lazier.&lt;/p&gt;

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

&lt;p&gt;[Add your video or GIF link here. A 20-second terminal recording of the dry run followed by &lt;code&gt;--apply&lt;/code&gt; works well.]&lt;/p&gt;

&lt;p&gt;Example of the dry-run output:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Screenshot 2026-10-12 at 9.41.03.png  -&amp;gt;  Images/dashboard-error-message.png
invoice_final_v2.pdf                  -&amp;gt;  Receipts/invoice-final-v2.pdf
notes.txt                             -&amp;gt;  Documents/meeting-notes-q4-planning.txt

Dry run: 3 files. Re-run with --apply to do it.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



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

&lt;p&gt;GitHub repo link here &lt;a href="https://dev.tourl"&gt;https://github.com/vikasfine17/sloth-sorter&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It's a single Python file using only the standard library, plus a README.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; an open-weight model served locally through &lt;a href="https://ollama.com" rel="noopener noreferrer"&gt;Ollama&lt;/a&gt;. The default is &lt;code&gt;llama3.2:3b&lt;/code&gt;, and &lt;code&gt;--model&lt;/code&gt; accepts anything Ollama can run.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;How it works:&lt;/strong&gt; for each file, the script sends the name, size and a short text preview to the local model. The model returns JSON with a category and a short kebab-case name. The script validates the category against a fixed list, sanitizes the name, and plans the move.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No dependencies:&lt;/strong&gt; it talks to Ollama's local HTTP API with &lt;code&gt;urllib&lt;/code&gt;, so there's nothing to install beyond Ollama itself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Built to be trusted with your files:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;Dry run by default, and nothing moves without &lt;code&gt;--apply&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Every move is logged to &lt;code&gt;.sloth-undo.json&lt;/code&gt;, so &lt;code&gt;--undo&lt;/code&gt; restores it.&lt;/li&gt;
&lt;li&gt;It never overwrites files; name collisions get a numeric suffix.&lt;/li&gt;
&lt;li&gt;It skips files modified in the last 10 minutes, so it won't touch a half-finished download.&lt;/li&gt;
&lt;li&gt;If the model is unreachable, it falls back to sorting by file extension.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;A Downloads folder is full of sensitive things: tax forms, IDs, receipts, contracts. Sending file names and previews to a hosted API just to tidy them would defeat the purpose.&lt;/p&gt;

&lt;p&gt;Running an open-weight model locally gave me:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; file contents never leave the machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Offline use and zero cost:&lt;/strong&gt; no API key, no per-file charges, and it works without internet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control:&lt;/strong&gt; I can swap models with one flag, or edit the categories and system prompt to match my own filing system. A closed API wouldn't let me do both of those while keeping the data local.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>automation</category>
      <category>hacktoberfest</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Sloth Sorter</title>
      <dc:creator>Pushpraj Patel</dc:creator>
      <pubDate>Thu, 22 Oct 2026 18:51:45 +0000</pubDate>
      <link>https://dev.to/vilaspatel938/sloth-sorter-3a2p</link>
      <guid>https://dev.to/vilaspatel938/sloth-sorter-3a2p</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;&lt;strong&gt;Sloth Sorter&lt;/strong&gt; is a lazy file tidier. Point it at a messy folder (my Downloads folder was the victim) and it reads each file, decides where it belongs, gives it a clean descriptive name, and files it away. You review a plan and approve it with one flag.&lt;/p&gt;

&lt;p&gt;It saves the chore everyone puts off: dragging &lt;code&gt;Screenshot (47).png&lt;/code&gt;, &lt;code&gt;invoice_final_v2.pdf&lt;/code&gt; and &lt;code&gt;download(3).zip&lt;/code&gt; into folders and renaming them. Anyone with a cluttered Downloads folder gets to be a little lazier.&lt;/p&gt;

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

&lt;p&gt;[Add your video or GIF link here. A 20-second terminal recording of the dry run followed by &lt;code&gt;--apply&lt;/code&gt; works well.]&lt;/p&gt;

&lt;p&gt;Example of the dry-run output:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Screenshot 2026-10-12 at 9.41.03.png  -&amp;gt;  Images/dashboard-error-message.png
invoice_final_v2.pdf                  -&amp;gt;  Receipts/invoice-final-v2.pdf
notes.txt                             -&amp;gt;  Documents/meeting-notes-q4-planning.txt

Dry run: 3 files. Re-run with --apply to do it.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



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

&lt;p&gt;GitHub repo link here. &lt;a href="https://github.com/vikasfine17/sloth-sorter" rel="noopener noreferrer"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It's a single Python file using only the standard library, plus a README.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; an open-weight model served locally through &lt;a href="https://ollama.com" rel="noopener noreferrer"&gt;Ollama&lt;/a&gt;. The default is &lt;code&gt;llama3.2:3b&lt;/code&gt;, and &lt;code&gt;--model&lt;/code&gt; accepts anything Ollama can run.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;How it works:&lt;/strong&gt; for each file, the script sends the name, size and a short text preview to the local model. The model returns JSON with a category and a short kebab-case name. The script validates the category against a fixed list, sanitizes the name, and plans the move.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No dependencies:&lt;/strong&gt; it talks to Ollama's local HTTP API with &lt;code&gt;urllib&lt;/code&gt;, so there's nothing to install beyond Ollama itself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Built to be trusted with your files:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;Dry run by default, and nothing moves without &lt;code&gt;--apply&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Every move is logged to &lt;code&gt;.sloth-undo.json&lt;/code&gt;, so &lt;code&gt;--undo&lt;/code&gt; restores it.&lt;/li&gt;
&lt;li&gt;It never overwrites files; name collisions get a numeric suffix.&lt;/li&gt;
&lt;li&gt;It skips files modified in the last 10 minutes, so it won't touch a half-finished download.&lt;/li&gt;
&lt;li&gt;If the model is unreachable, it falls back to sorting by file extension.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;A Downloads folder is full of sensitive things: tax forms, IDs, receipts, contracts. Sending file names and previews to a hosted API just to tidy them would defeat the purpose.&lt;/p&gt;

&lt;p&gt;Running an open-weight model locally gave me:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; file contents never leave the machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Offline use and zero cost:&lt;/strong&gt; no API key, no per-file charges, and it works without internet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control:&lt;/strong&gt; I can swap models with one flag, or edit the categories and system prompt to match my own filing system. A closed API wouldn't let me do both of those while keeping the data local.&lt;/li&gt;
&lt;/ul&gt;

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