<?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: Amit Rawal</title>
    <description>The latest articles on DEV Community by Amit Rawal (@amit_rawal_93de441d25a95a).</description>
    <link>https://dev.to/amit_rawal_93de441d25a95a</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%2F4173187%2F1936c5b7-e084-4f03-8d09-24f405a9eea9.jpg</url>
      <title>DEV Community: Amit Rawal</title>
      <link>https://dev.to/amit_rawal_93de441d25a95a</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/amit_rawal_93de441d25a95a"/>
    <language>en</language>
    <item>
      <title>HauntedShell</title>
      <dc:creator>Amit Rawal</dc:creator>
      <pubDate>Wed, 28 Oct 2026 15:09:33 +0000</pubDate>
      <link>https://dev.to/amit_rawal_93de441d25a95a/hauntedshell-d9o</link>
      <guid>https://dev.to/amit_rawal_93de441d25a95a/hauntedshell-d9o</guid>
      <description>&lt;p&gt;A terminal that a ghost has moved into.&lt;br&gt;
Run normal shell commands. After each one, a ghost (a small open-weight model running locally with Ollama) whispers a short, creepy comment about what you did.&lt;br&gt;
No internet, no API key, no cloud. One Python file using only the standard library.&lt;br&gt;
Run it&lt;/p&gt;

&lt;h1&gt;
  
  
  1. Start Ollama and download the ghost's model (one time)
&lt;/h1&gt;

&lt;p&gt;ollama serve&lt;br&gt;
ollama pull qwen2.5:3b&lt;/p&gt;

&lt;h1&gt;
  
  
  2. Enter the haunted shell
&lt;/h1&gt;

&lt;p&gt;python3 haunted.py&lt;br&gt;
​&lt;br&gt;
Type exit to leave.&lt;br&gt;
Options&lt;br&gt;
Environment variable&lt;br&gt;
Default&lt;br&gt;
What it does&lt;br&gt;
GHOST_MODEL&lt;br&gt;
qwen2.5:3b&lt;br&gt;
Which Ollama model plays the ghost&lt;br&gt;
OLLAMA_URL&lt;br&gt;
&lt;a href="http://localhost:11434" rel="noopener noreferrer"&gt;http://localhost:11434&lt;/a&gt;&lt;br&gt;
Where Ollama is running&lt;br&gt;
Example: GHOST_MODEL=llama3.2:3b python3 haunted.py&lt;br&gt;
If Ollama isn't running, the ghost uses built-in backup lines, so the shell still works.&lt;br&gt;
Note&lt;br&gt;
Commands really run on your machine, just like in your normal terminal. The ghost only comments; it never runs anything by itself.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>hackathon</category>
      <category>ai</category>
    </item>
    <item>
      <title>🦥 SlothSort</title>
      <dc:creator>Amit Rawal</dc:creator>
      <pubDate>Sat, 24 Oct 2026 16:45:58 +0000</pubDate>
      <link>https://dev.to/amit_rawal_93de441d25a95a/slothsort-2f1l</link>
      <guid>https://dev.to/amit_rawal_93de441d25a95a/slothsort-2f1l</guid>
      <description>&lt;p&gt;Tidies your messy folder while you nap. Runs 100% on your laptop.&lt;br&gt;
SlothSort looks at the loose files in a folder (your Desktop, Downloads, anywhere), reads what is inside them, and suggests a clear name and the right folder for each one. You review the plan, change anything you like, and press one button. Every run can be undone.&lt;br&gt;
It uses small open-weight models running locally through Ollama. No cloud, no API key, no account. Your resumes, invoices and screenshots never leave your machine.&lt;br&gt;
Built for the DEV "Slow and Steady" challenge (International Sloth Day 2026).&lt;br&gt;
How it works&lt;br&gt;
 folder ──► scanner ──► planner ──► review in browser ──► mover ──► history (undo)&lt;br&gt;
             │            │&lt;br&gt;
             │            ├─ videos / installers / zips → sorted by type (no AI needed)&lt;br&gt;
             │            ├─ images → moondream describes them → qwen2.5 decides&lt;br&gt;
             │            └─ PDFs / docx / text / code → qwen2.5 reads the start and decides&lt;br&gt;
             └─ pulls a short text sample from each file (pypdf, docx XML, plain text)&lt;br&gt;
​&lt;br&gt;
Part&lt;br&gt;
File&lt;br&gt;
What it does&lt;br&gt;
Scanner&lt;br&gt;
slothsort/scanner.py&lt;br&gt;
Lists loose files and reads a ~1500-character sample from each one&lt;br&gt;
Brain&lt;br&gt;
slothsort/brain.py&lt;br&gt;
Talks to Ollama; describes images; asks for {folder, new_name, reason} as JSON&lt;br&gt;
Planner&lt;br&gt;
slothsort/planner.py&lt;br&gt;
Builds the plan file by file, with a cache so unchanged files are never re-analysed&lt;br&gt;
Mover&lt;br&gt;
slothsort/mover.py&lt;br&gt;
Moves approved files, never overwrites, never deletes, logs every run for undo&lt;br&gt;
Web app&lt;br&gt;
slothsort/app.py, static/index.html&lt;br&gt;
FastAPI server and a single page that streams the plan live&lt;br&gt;
Making a 3B model reliable&lt;br&gt;
Small models can drift, so SlothSort keeps them on a short leash:&lt;br&gt;
Structured output. Ollama's format option forces the reply to match a JSON schema.&lt;br&gt;
Folder is an enum. The model can only choose from a fixed list of folders (plus folders that already exist), so it cannot invent Stuff2/.&lt;br&gt;
Temperature 0. The same file gets the same suggestion every time.&lt;br&gt;
Skip the model when it can't help. A .mov or .dmg has no readable content, so a simple rule sorts it instantly.&lt;br&gt;
Two small models instead of one big one. moondream turns a screenshot into a sentence, then qwen2.5:3b makes the decision. Both fit in 8 GB of RAM.&lt;br&gt;
Safety&lt;br&gt;
Nothing moves until you click Move files.&lt;br&gt;
Files are only moved, never deleted, and never overwritten (Report (2).pdf).&lt;br&gt;
Only files sitting directly in the chosen folder can be moved.&lt;br&gt;
Every run is saved in data/history/ and can be undone from the page.&lt;br&gt;
The server only listens on 127.0.0.1, and API calls need a custom header so other websites can't trigger moves.&lt;br&gt;
Run it&lt;br&gt;
Needs Python 3.10+ and Ollama.&lt;/p&gt;

&lt;h1&gt;
  
  
  1. Get the models (one time, ~3.5 GB)
&lt;/h1&gt;

&lt;p&gt;ollama serve                         # keep this running in its own tab&lt;br&gt;
ollama pull qwen2.5:3b&lt;br&gt;
ollama pull moondream&lt;/p&gt;

&lt;h1&gt;
  
  
  2. Install and start SlothSort
&lt;/h1&gt;

&lt;p&gt;git clone &lt;a href="https://github.com/Amitrawal1/SlothSort.git" rel="noopener noreferrer"&gt;https://github.com/Amitrawal1/SlothSort.git&lt;/a&gt;&lt;br&gt;
cd SlothSort&lt;br&gt;
pip install -r requirements.txt&lt;br&gt;
python3 demo/make_messy_folder.py    # optional: a fake messy folder to play with&lt;br&gt;
python3 -m slothsort.app             # open &lt;a href="http://127.0.0.1:8000" rel="noopener noreferrer"&gt;http://127.0.0.1:8000&lt;/a&gt;&lt;br&gt;
​&lt;br&gt;
Command line&lt;br&gt;
python3 -m slothsort.cli plan  demo/messy     # show the plan only&lt;br&gt;
python3 -m slothsort.cli apply demo/messy     # show the plan, ask, then move&lt;br&gt;
python3 -m slothsort.cli runs                 # list past runs&lt;br&gt;
python3 -m slothsort.cli undo         # put files back&lt;br&gt;
​&lt;br&gt;
Swap the models&lt;br&gt;
Any Ollama model works. Bigger models give better names; smaller ones are faster.&lt;br&gt;
SLOTHSORT_TEXT_MODEL=llama3.2:3b SLOTHSORT_VISION_MODEL=qwen2.5vl:3b python3 -m slothsort.app&lt;br&gt;
​&lt;br&gt;
You can also edit the folder list in slothsort/config.py.&lt;br&gt;
Project layout&lt;br&gt;
SlothSort/&lt;br&gt;
├── slothsort/&lt;br&gt;
│   ├── config.py      settings, models, folder list&lt;br&gt;
│   ├── scanner.py     read files&lt;br&gt;
│   ├── brain.py       talk to Ollama&lt;br&gt;
│   ├── planner.py     build the plan (+ cache)&lt;br&gt;
│   ├── mover.py       move + undo&lt;br&gt;
│   ├── app.py         web server&lt;br&gt;
│   └── cli.py         command line&lt;br&gt;
├── static/index.html  the web page&lt;br&gt;
├── demo/make_messy_folder.py&lt;br&gt;
└── requirements.txt&lt;br&gt;
​&lt;br&gt;
License&lt;br&gt;
MIT&lt;/p&gt;

</description>
      <category>ai</category>
      <category>programming</category>
      <category>hf26challenge</category>
      <category>devchallenge</category>
    </item>
    <item>
      <title>ListSmart: Turn Product Photos into Meesho/Etsy Listings Locally with AI</title>
      <dc:creator>Amit Rawal</dc:creator>
      <pubDate>Sun, 18 Oct 2026 09:48:12 +0000</pubDate>
      <link>https://dev.to/amit_rawal_93de441d25a95a/listsmart-turn-product-photos-into-meeshoetsy-listings-locally-with-ai-6p3</link>
      <guid>https://dev.to/amit_rawal_93de441d25a95a/listsmart-turn-product-photos-into-meeshoetsy-listings-locally-with-ai-6p3</guid>
      <description>&lt;h1&gt;
  
  
  ListSmart: Turn Product Photos into E-commerce Listings with AI
&lt;/h1&gt;

&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge — Week 2: Side Hustle&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;ListSmart&lt;/strong&gt; converts product photos into ready-to-publish Meesho and Etsy listings, including titles, descriptions, five bullet points, search tags, and suggested prices.&lt;/p&gt;

&lt;p&gt;It helps small sellers save time, write better listings, and price products more confidently.&lt;/p&gt;

&lt;p&gt;Key features:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;📸 Product detail extraction using Gemma.&lt;/li&gt;
&lt;li&gt;✍️ Listing generation using fine-tuned Qwen3-8B.&lt;/li&gt;
&lt;li&gt;💰 Competitor-based pricing using SerpApi.&lt;/li&gt;
&lt;li&gt;🔒 Local inference for privacy and zero per-listing API costs.&lt;/li&gt;
&lt;li&gt;✏️ Editable listings with character counters and one-click copy.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The proposed business model is 20 free listings, followed by approximately ₹199/month. Pricing is still being validated.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo and Code
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;GitHub:&lt;/strong&gt; &lt;a href="https://github.com/Amitrawal1/ListSmart" rel="noopener noreferrer"&gt;https://github.com/Amitrawal1/ListSmart&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Run locally:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/Amitrawal1/ListSmart.git
&lt;span class="nb"&gt;cd &lt;/span&gt;ListSmart/backend
python3 &lt;span class="nt"&gt;-m&lt;/span&gt; venv .venv
.venv/bin/pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; requirements.txt
.venv/bin/uvicorn app.main:app &lt;span class="nt"&gt;--port&lt;/span&gt; 8001
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open &lt;code&gt;http://localhost:8001&lt;/code&gt;. Add &lt;code&gt;?demo=1&lt;/code&gt; to try sample data.&lt;/p&gt;

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

&lt;p&gt;The project uses Gemma for vision, Qwen3-8B with LoRA fine-tuning for listing generation, Tinker for training, and SerpApi for competitor pricing.&lt;/p&gt;

&lt;p&gt;The stack includes FastAPI, Ollama, and a lightweight HTML/CSS/JavaScript frontend. The repository also includes training, evaluation, and setup scripts.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Learned from Version 1
&lt;/h2&gt;

&lt;p&gt;My first fine-tune reduced validation loss from &lt;strong&gt;2.61 to 1.48&lt;/strong&gt;, but performed worse than the base model on important quality metrics.&lt;/p&gt;

&lt;p&gt;The blind evaluation preferred the base model &lt;strong&gt;77.5%&lt;/strong&gt; of the time, compared with just &lt;strong&gt;0.8%&lt;/strong&gt; for v1.&lt;/p&gt;

&lt;p&gt;The main problems were repetitive keyword stuffing and invented product details.&lt;/p&gt;

&lt;p&gt;For v2, I'm cleaning the training data, removing unsupported claims, and enforcing platform rules. The final evaluation is still pending, and I'll share the results once measured.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Open Source Matters
&lt;/h2&gt;

&lt;p&gt;Open models let me inspect training data, fix model behaviour, and run inference locally. Sellers can keep product photos on their own devices, while local inference avoids per-listing API charges.&lt;/p&gt;

&lt;p&gt;The project also reduces dependence on a single model provider and makes the system easier to modify.&lt;/p&gt;

&lt;h2&gt;
  
  
  What's Next?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Complete v2 training and evaluation.&lt;/li&gt;
&lt;li&gt;Deploy a public demo.&lt;/li&gt;
&lt;li&gt;Collect feedback from real sellers.&lt;/li&gt;
&lt;li&gt;Validate the subscription pricing.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Technologies and Challenge Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Gemma:&lt;/strong&gt; Product-photo understanding.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tinker:&lt;/strong&gt; LoRA fine-tuning and model evaluation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;SerpApi:&lt;/strong&gt; Competitor-price discovery.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Repository:&lt;/strong&gt; &lt;a href="https://github.com/Amitrawal1/ListSmart" rel="noopener noreferrer"&gt;https://github.com/Amitrawal1/ListSmart&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Thanks for reading!&lt;/p&gt;

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