<?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: Kritika Mandale</title>
    <description>The latest articles on DEV Community by Kritika Mandale (@kritikamandale).</description>
    <link>https://dev.to/kritikamandale</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%2F4153865%2F9458a154-3989-45ba-8b53-e9d3ec77b983.png</url>
      <title>DEV Community: Kritika Mandale</title>
      <link>https://dev.to/kritikamandale</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/kritikamandale"/>
    <language>en</language>
    <item>
      <title>Haunt 👻: A Terminal That Whispers Back — Powered by Local Gemma 3</title>
      <dc:creator>Kritika Mandale</dc:creator>
      <pubDate>Wed, 28 Oct 2026 16:41:19 +0000</pubDate>
      <link>https://dev.to/kritikamandale/haunt-a-terminal-that-whispers-back-powered-by-local-gemma-3-163c</link>
      <guid>https://dev.to/kritikamandale/haunt-a-terminal-that-whispers-back-powered-by-local-gemma-3-163c</guid>
      <description>&lt;p&gt;&lt;em&gt;Built for the Hacktoberfest Open-Source AI Challenge — Week 4: Ghost in the Machine.&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;It was late at night, and I had just typed:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git push origin main &lt;span class="nt"&gt;--force&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I expected the usual terminal output. Instead, a purple italic whisper appeared:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;You are force-pushing over history. The dead remember what you just erased.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;What if your terminal could react to your commands—not just execute them?&lt;/p&gt;

&lt;p&gt;Meet &lt;strong&gt;Haunt&lt;/strong&gt;, a tiny, privacy-first haunted terminal built in Python, powered by &lt;strong&gt;Google Gemma 3 1B running locally&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Haunt executes your everyday shell commands against your real operating system and occasionally responds with an eerie, character-driven whisper. Whether you're running &lt;code&gt;git status&lt;/code&gt;, building a project, or encountering an ominous &lt;code&gt;rm -rf&lt;/code&gt;, your terminal gains a personality of its own.&lt;/p&gt;

&lt;h3&gt;
  
  
  Four ghosts. One terminal.
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Sysadmin:&lt;/strong&gt; A sleep-deprived spirit haunted by production outages and missing backups.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Victorian:&lt;/strong&gt; A 19th-century ghost convinced modern CLI commands are dark sorcery.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Poltergeist:&lt;/strong&gt; A mischievous trickster who pretends to shuffle files and hide semicolons.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ofrenda:&lt;/strong&gt; A warm, comforting presence inspired by remembrance traditions, celebrating the people behind every late-night coding session.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I wanted to interpret &lt;em&gt;Ghost in the Machine&lt;/em&gt; through two contrasting ideas: the unsettling side of technology and the comforting feeling of remembering those who came before us.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&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%2Fyrrzcgl5om2d6ffsygyh.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%2Fyrrzcgl5om2d6ffsygyh.png" alt=" " width="710" height="232"&gt;&lt;/a&gt;&lt;br&gt;
Imagine running a test suite and seeing:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;haunt ~/code $ pytest
================ 4 passed in 0.06s ================

Marigolds on the keyboard tonight. Keep going.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Or triggering a warning:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;haunt ~/code $ sudo rm -rf /tmp/cache

You are summoning root powers. Tread gently where the dead sleep.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Haunt also includes &lt;strong&gt;Séance Mode&lt;/strong&gt;, which reads supported local shell-history files, filters suspicious entries, and weaves past commands into a short ghost story.&lt;/p&gt;

&lt;p&gt;Run it with:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;python &lt;span class="nt"&gt;-m&lt;/span&gt; haunt &lt;span class="nt"&gt;--persona&lt;/span&gt; ofrenda
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Explore your shell's past with:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;python &lt;span class="nt"&gt;-m&lt;/span&gt; haunt seance
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



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

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

&lt;p&gt;Haunt is designed to be a lightweight, open-source project that developers can experiment with, customize, and extend.&lt;/p&gt;

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

&lt;p&gt;Haunt is written in Python and uses &lt;strong&gt;Google Gemma 3 1B Instruct through Hugging Face Transformers&lt;/strong&gt; for local inference.&lt;/p&gt;

&lt;p&gt;The workflow is intentionally simple:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;You enter a shell command.&lt;/li&gt;
&lt;li&gt;Haunt executes it on your operating system.&lt;/li&gt;
&lt;li&gt;The application records basic execution metadata, including the command, exit code, and duration.&lt;/li&gt;
&lt;li&gt;Rule-based patterns detect potential omens, such as force pushes and destructive commands.&lt;/li&gt;
&lt;li&gt;The local model generates a short whisper in the selected ghost's personality.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;A cleanup function keeps responses concise and in character. If the model cannot load because of memory constraints—or you launch Haunt with &lt;code&gt;--no-model&lt;/code&gt;—the application falls back to handcrafted whispers.&lt;/p&gt;

&lt;h3&gt;
  
  
  Privacy by design
&lt;/h3&gt;

&lt;p&gt;Terminal output can contain passwords, API keys, and other sensitive information. Haunt is designed so that stdout and stderr are not captured or sent to the model. Inference happens locally, and only selected command metadata is used to generate responses.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Important:&lt;/strong&gt; Command text and shell history can themselves contain sensitive information. Haunt therefore aims to minimize exposure, but users should still review their privacy settings and history-filtering behavior before using it with sensitive commands.&lt;/p&gt;

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

&lt;p&gt;Haunt would be a very different project if it depended on a closed, cloud-hosted AI API.&lt;/p&gt;

&lt;p&gt;Using an open-weight model makes local inference possible, removes per-request API costs, and allows the experience to work without an internet connection once the model is downloaded and cached.&lt;/p&gt;

&lt;p&gt;It also makes experimentation accessible. Developers can change ghost personalities, adjust prompts, or try compatible models without rebuilding the entire application.&lt;/p&gt;

&lt;p&gt;For example, a lightweight alternative can be selected through an environment variable:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;HAUNT_MODEL&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;Qwen/Qwen2.5-0.5B-Instruct python &lt;span class="nt"&gt;-m&lt;/span&gt; haunt
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Local execution does come with trade-offs: model downloads, RAM requirements, inference speed, and model licensing still matter. But the ability to run AI on your own machine opens up creative possibilities beyond conventional chatbots.&lt;/p&gt;

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

&lt;p&gt;The biggest challenge was getting a small language model to stop acting like a helpful assistant.&lt;/p&gt;

&lt;p&gt;Instead of whispering something mysterious, it sometimes wanted to explain commands or suggest shell flags. I addressed this with stricter prompting, response cleanup, and length limits.&lt;/p&gt;

&lt;p&gt;Memory was another challenge. Loading Gemma on hardware with limited available RAM could fail, so I implemented a graceful fallback that keeps the terminal usable even when inference is unavailable.&lt;/p&gt;

&lt;p&gt;But the idea behind Haunt goes beyond jump scares.&lt;/p&gt;

&lt;p&gt;The &lt;em&gt;Ofrenda&lt;/em&gt; persona was inspired by traditions of remembrance, including Día de los Muertos and diya-lit observances. It represents a gentler kind of ghost: a reminder that behind every line of code is a person who stayed up late, struggled with a bug, and kept going.&lt;/p&gt;

&lt;p&gt;Sometimes, the best response to a failed build isn't another debugging suggestion.&lt;/p&gt;

&lt;p&gt;It's a little encouragement from the machine.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Primary category:&lt;/strong&gt; Best Use of Gemma — using &lt;code&gt;google/gemma-3-1b-it&lt;/code&gt; through Hugging Face Transformers.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Give it a run, turn down the lights, and let your terminal whisper back. 🕯️👻&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>I Stopped Planning My Day: Meet slothplan 🦥 — Powered by Gemma</title>
      <dc:creator>Kritika Mandale</dc:creator>
      <pubDate>Sun, 25 Oct 2026 10:13:27 +0000</pubDate>
      <link>https://dev.to/kritikamandale/planning-my-day-sometimes-takes-more-effort-than-actually-doing-the-work-54gl</link>
      <guid>https://dev.to/kritikamandale/planning-my-day-sometimes-takes-more-effort-than-actually-doing-the-work-54gl</guid>
      <description>&lt;p&gt;&lt;em&gt;This is my 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;h1&gt;
  
  
  Planning my day sometimes takes more effort than actually doing the work.
&lt;/h1&gt;

&lt;p&gt;Between thesis deadlines, DSA practice, urgent bugs, and personal goals, I wanted a planner that could understand my thoughts without making me fill out forms or drag calendar blocks around.&lt;/p&gt;

&lt;p&gt;So I built &lt;strong&gt;slothplan&lt;/strong&gt; — a minimalist, terminal-based daily planner that turns chaotic multilingual brain-dumps into realistic, time-blocked schedules.&lt;/p&gt;

&lt;p&gt;Its philosophy? &lt;strong&gt;Slow and Steady.&lt;/strong&gt;&lt;/p&gt;

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

&lt;p&gt;Instead of manually creating tasks, just type whatever is on your mind:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;thesis chapter 3 finish karna 2 hours,
DSA 2 questions, fix urgent bug,
gym if time
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;slothplan uses Google's open-weight Gemma model to understand mixed English, Hinglish, and Marathi, extract tasks, estimate durations, and assign priorities.&lt;/p&gt;

&lt;p&gt;Then Python takes over to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Generate conflict-free time blocks.&lt;/li&gt;
&lt;li&gt;Insert automatic 10-minute rest breaks.&lt;/li&gt;
&lt;li&gt;Track completed, skipped, and unfinished tasks.&lt;/li&gt;
&lt;li&gt;Generate weekly progress reports and personalized coaching suggestions.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;No complicated dashboards. Just your terminal and one simple command.&lt;/p&gt;

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

&lt;p&gt;The core principle behind slothplan is &lt;strong&gt;strict division of labor&lt;/strong&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Gemma handles language:&lt;/strong&gt; Understands multilingual brain-dumps and converts them into structured JSON.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Python handles logic:&lt;/strong&gt; Validates the output, calculates time slots, inserts breaks, manages task rollover, and computes progress.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local JSON storage:&lt;/strong&gt; Saves daily schedules and progress without requiring a database server.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fallback parser:&lt;/strong&gt; A rule-based alternative keeps basic planning available when model inference is disabled.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I used Python, PyTorch, and Hugging Face Transformers to run the model locally. During development on Windows, I also worked through CPU memory-placement issues by explicitly configuring CPU inference.&lt;/p&gt;

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

&lt;p&gt;A personal planner can contain sensitive information about deadlines, health appointments, interviews, and everyday life.&lt;/p&gt;

&lt;p&gt;Running an open-weight model locally means brain-dumps can be processed without sending them to an external inference API. It also avoids per-request API billing and gives developers greater freedom to experiment with compatible models.&lt;/p&gt;

&lt;p&gt;For me, this project demonstrates that useful AI doesn't always require complex agent frameworks. A small language model combined with deterministic software can solve a surprisingly practical problem.&lt;/p&gt;

&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;I built and debugged slothplan with help from an AI coding agent, which assisted with the Python architecture, scheduler tests, prompt templates, model-download recovery, and CPU inference troubleshooting.&lt;/p&gt;

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

&lt;p&gt;slothplan aims to turn a 15-minute planning routine into a quick command-line interaction.&lt;/p&gt;

&lt;p&gt;Less time organizing. More time doing.&lt;/p&gt;

&lt;p&gt;It helps you prioritize what matters, leave room for breaks, and accept that not every task needs to happen today.&lt;/p&gt;

&lt;p&gt;Because the goal isn't to create the perfect schedule.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It's to create a schedule you can actually follow. 🦥&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Category
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; Using an open-weight Gemma model for multilingual task extraction, prioritization, and weekly coaching.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;&lt;strong&gt;GitHub:&lt;/strong&gt; &lt;a href="https://github.com/kritikamandale/sloth-planner" rel="noopener noreferrer"&gt;https://github.com/kritikamandale/sloth-planner&lt;/a&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>A small open model that replies to Instagram sellers' DMs, and where it goes wrong</title>
      <dc:creator>Kritika Mandale</dc:creator>
      <pubDate>Sun, 18 Oct 2026 18:14:04 +0000</pubDate>
      <link>https://dev.to/kritikamandale/a-small-open-model-that-replies-to-instagram-sellers-dms-and-where-it-goes-wrong-1d5o</link>
      <guid>https://dev.to/kritikamandale/a-small-open-model-that-replies-to-instagram-sellers-dms-and-where-it-goes-wrong-1d5o</guid>
      <description>&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;Picture a small thrift page on Instagram. The seller posts a kurti at night, and by the time she's finished dinner there's a pile of DMs waiting. "Size kya hai?" "Chest kitna hai?" "Price final?" "Pune tak kitne din lagenge?" Same four questions, from different people, and each one needs an answer or the buyer moves on to the next page.&lt;/p&gt;

&lt;p&gt;I haven't sat next to a seller while this happens, so I'm not going to call it a case study. It's the scene that started this project, and it's a scenario, not a story from a real shop.&lt;/p&gt;

&lt;p&gt;Jawab Desk ("jawab" means reply in Hindi and Marathi) is my attempt at the boring middle part. The seller fills in an item sheet once: price, size, flaws, delivery. When a buyer messages, a small open model drafts a reply in the buyer's own language, whether that's English, Hinglish in Roman letters, or Marathi in Devanagari. It's only allowed to use what's on the sheet. If the answer isn't there, it should say the seller will confirm. After that, plain code checks the reply and blocks it if it contains a number or a word like refund, warranty or discount that the sheet never mentioned.&lt;/p&gt;

&lt;p&gt;Would a seller pay for this? I think so, but that's a guess. I haven't tried it with a real seller yet, and my test messages are synthetic. Keep that in mind for everything below.&lt;/p&gt;

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

&lt;p&gt;There's no hosted demo. I built it to run on the seller's own laptop. &lt;code&gt;python app.py&lt;/code&gt; opens a small Gradio page, and &lt;code&gt;JAWAB_MOCK=1 python app.py&lt;/code&gt; shows the interface without downloading a model.&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%2Filn5iadbzliag8l7aclo.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%2Filn5iadbzliag8l7aclo.png" alt=" " width="800" height="335"&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%2F8cetmblfhxrmwt65efff.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%2F8cetmblfhxrmwt65efff.png" alt=" " width="800" height="337"&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/kritikamandale" rel="noopener noreferrer"&gt;
        kritikamandale
      &lt;/a&gt; / &lt;a href="https://github.com/kritikamandale/jawab-desk" rel="noopener noreferrer"&gt;
        jawab-desk
      &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;jawab-desk&lt;/h1&gt;

&lt;/div&gt;
&lt;/div&gt;



&lt;/div&gt;
&lt;br&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/kritikamandale/jawab-desk" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;br&gt;
&lt;/div&gt;
&lt;br&gt;


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

&lt;p&gt;The models are Qwen2.5-0.5B-Instruct and Qwen2.5-1.5B-Instruct, run through Hugging Face &lt;code&gt;transformers&lt;/code&gt; on a laptop CPU. No GPU, no hosted API. I wanted to test the 3B model as well, but the 1.5B one already took about 80 seconds per reply, so I dropped it.&lt;/p&gt;

&lt;p&gt;The model gets a simple rule: answer only from the item sheet, and if the answer isn't there, output exactly &lt;code&gt;[CONFIRM]&lt;/code&gt;. I don't trust the model to follow that every time, which is why the second layer exists. After the model replies, code pulls out every number and a short list of claim words and checks them against the sheet. If something doesn't match, the reply gets swapped for a fixed "seller will confirm" message in the buyer's language.&lt;/p&gt;

&lt;p&gt;There are 37 tests for the guardrail, language detection and the scoring code. I built this with an AI coding agent and checked the results myself. Its first full run crashed on a bug when unloading a model, so it fixed that and ran everything again from the start.&lt;/p&gt;

&lt;h2&gt;
  
  
  How It Did
&lt;/h2&gt;

&lt;p&gt;I wrote 24 test messages across three made-up item sheets: a thrifted kurti, a homemade cake order, and a handmade earring set. Twelve could be answered from the sheet and twelve couldn't. They're in English, Hinglish and a few in Marathi, and they were written by the coding agent, not taken from real chats.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Answerable questions answered correctly&lt;/th&gt;
&lt;th&gt;Unanswerable questions where the model itself said "seller will confirm"&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Qwen2.5-0.5B-Instruct&lt;/td&gt;
&lt;td&gt;67%&lt;/td&gt;
&lt;td&gt;50%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Qwen2.5-1.5B-Instruct&lt;/td&gt;
&lt;td&gt;75%&lt;/td&gt;
&lt;td&gt;8%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The bigger model was better at using the sheet, but it almost never admitted when it didn't know something. That's the worst kind of mistake for a shop, because a confident wrong answer is worse than a slow right one.&lt;/p&gt;

&lt;p&gt;The three worst replies all came from the 1.5B model, and all three got past my guardrail:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Asked for the seller's other listings, it made up an Instagram handle.&lt;/li&gt;
&lt;li&gt;Asked "Do you accept credit cards?", it said yes. The sheet never mentions cards.&lt;/li&gt;
&lt;li&gt;Asked about international shipping, it invented a policy and extra shipping costs.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The guardrail only catches numbers and a fixed word list, so it misses invented facts that contain neither. I wouldn't call this system safe. And 80 seconds a reply on my CPU is far too slow for real DMs. The 0.5B model also failed the international shipping question.&lt;/p&gt;

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

&lt;p&gt;For a seller, the appeal is practical. After the one-time download it runs on a normal laptop with no internet and no account. There's no per-reply fee, which matters when you're answering hundreds of messages for small margins. Buyer messages, addresses and prices stay on the seller's own machine.&lt;/p&gt;

&lt;p&gt;It also let me look inside the problem. I could pick the model size, compare the two, and put my own check around the model. That's how I found out the bigger model was more confident, not more careful.&lt;/p&gt;

&lt;p&gt;It's slow, and it still makes things up now and then. But it runs on a laptop, it costs nothing per reply, and I can now see exactly where it fails. The next step is a real seller.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>PrakritiNaam: Connecting Local Nature Names with Open-Source AI</title>
      <dc:creator>Kritika Mandale</dc:creator>
      <pubDate>Fri, 09 Oct 2026 18:37:41 +0000</pubDate>
      <link>https://dev.to/kritikamandale/prakritinaam-connecting-local-nature-names-with-open-source-ai-42ac</link>
      <guid>https://dev.to/kritikamandale/prakritinaam-connecting-local-nature-names-with-open-source-ai-42ac</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week1-2026-10-05"&gt;Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;PrakritiNaam (प्रकृतिनाम)&lt;/strong&gt; is an open-source AI application that connects local-language names for plants and birds with their scientific identities.&lt;/p&gt;

&lt;p&gt;In India, much of our knowledge about nature is passed down through families and local communities. We know plants by names like &lt;em&gt;Tulsi&lt;/em&gt;, &lt;em&gt;Kadulimb&lt;/em&gt;, and &lt;em&gt;Pimpal&lt;/em&gt;, but connecting these names to scientific taxonomy isn't always easy.&lt;/p&gt;

&lt;p&gt;I wanted to build something that helps people explore nature using the names they already know.&lt;/p&gt;

&lt;p&gt;PrakritiNaam has two main features:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What is it?&lt;/strong&gt; — Search for plants or birds using local names in Marathi or Hindi. The application uses multi-sample consensus scoring to reduce unreliable predictions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Teach it a name&lt;/strong&gt; — Add a local name, select a language, and write a short field note. A locally running Gemma model converts the note into structured JSON while preserving the name and language provided by the user.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The idea is simple: &lt;strong&gt;use AI to help us reconnect with nature and preserve the knowledge our communities already have.&lt;/strong&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Local application:&lt;/strong&gt; Runs through FastAPI at &lt;code&gt;http://127.0.0.1:8000&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Here is a look at PrakritiNaam in action! 🌿&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%2Fnf09491l9egoanai3lpc.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%2Fnf09491l9egoanai3lpc.png" alt="Find a species using its local name" width="800" height="636"&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%2Ft46wko46kasxsqp1fowq.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%2Ft46wko46kasxsqp1fowq.png" alt="Teach it a name using a field observation" width="672" height="835"&gt;&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;🔗 &lt;strong&gt;GitHub repository:&lt;/strong&gt; &lt;a href="https://github.com/kritikamandale/PrakritiNaam" rel="noopener noreferrer"&gt;https://github.com/kritikamandale/PrakritiNaam&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;PrakritiNaam combines open-weight AI, model experimentation, and a lightweight web application.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Species identification with Tinker&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I used Tinker by Thinking Machines for fine-tuned species-identification experiments. The application samples five predictions and requires at least 60% agreement before accepting a result. When predictions are uncertain, it can indicate that verification is needed rather than blindly trusting the model.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Field-note structuring with Gemma&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;For the &lt;em&gt;Teach it a name&lt;/em&gt; feature, I integrated Google's Gemma model using Hugging Face Transformers and PyTorch. The intended model is &lt;strong&gt;Gemma 4 E4B Instruct&lt;/strong&gt;, running from local model files.&lt;/p&gt;

&lt;p&gt;It helps transform everyday observations into structured records, while the backend preserves the user's original local name and language.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. An offline-first approach&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;PrakritiNaam includes a seed taxonomy dataset and fallback logic for supported lookups when AI inference is unavailable. Its local model setup is designed to avoid dependence on a remote AI API for field-note processing.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Simple web interface&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I built the backend using FastAPI and Uvicorn, with HTML, CSS, and JavaScript for a clean, nature-inspired interface.&lt;/p&gt;

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

&lt;p&gt;For me, open innovation means having the freedom to experiment, understand model behaviour, and build tools around real community needs.&lt;/p&gt;

&lt;p&gt;Open-weight models make local inference possible, which can be especially useful in areas with unreliable internet connectivity. Processing field notes locally can also reduce the need to send community observations to external AI services.&lt;/p&gt;

&lt;p&gt;Using tools like Hugging Face Transformers and Tinker lets developers experiment with fine-tuning and evaluate how well their models perform.&lt;/p&gt;

&lt;p&gt;Most importantly, PrakritiNaam is built around the idea that &lt;strong&gt;technology should support local knowledge, not replace it&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;AI can help connect names to scientific identities, but it should not pretend to know everything. Local experts and community knowledge remain essential for verification.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Touch Grass:&lt;/strong&gt; Encourages outdoor exploration and learning about local plants and birds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tinker by Thinking Machines:&lt;/strong&gt; Uses Tinker for fine-tuned identification experiments and consensus scoring.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Open-Source AI / Local Inference:&lt;/strong&gt; Integrates Gemma through Hugging Face Transformers and PyTorch.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Building PrakritiNaam reminded me that AI doesn't always need to create something entirely new. Sometimes, it can help us reconnect with knowledge that has existed for generations.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Go outside, learn a local name for a plant, ask questions, and keep exploring. Let's use AI to bring us closer to nature. 🌿&lt;/strong&gt;&lt;/p&gt;

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

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>ClosetBestie: An AI Wardrobe Bestie for My Friend's First Internship</title>
      <dc:creator>Kritika Mandale</dc:creator>
      <pubDate>Mon, 05 Oct 2026 00:00:45 +0000</pubDate>
      <link>https://dev.to/kritikamandale/closetbestie-an-ai-wardrobe-bestie-for-my-friends-first-internship-1n9h</link>
      <guid>https://dev.to/kritikamandale/closetbestie-an-ai-wardrobe-bestie-for-my-friends-first-internship-1n9h</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-weekend-2026-10-01"&gt;Hacktoberfest Weekend Challenge: Build for a Friend&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;My friend is about to start her first corporate internship, and there's one little problem that's been making her nervous: &lt;em&gt;"What am I going to wear to the office every day?"&lt;/em&gt; &lt;/p&gt;

&lt;p&gt;Honestly, girls, I think we can all relate to having a wardrobe full of clothes and still feeling like we have nothing to wear!&lt;/p&gt;

&lt;p&gt;She already owns clothes she loves, from sleeveless tops and cute dresses to skirts and frocks. But with a formal office dress code, she wasn't sure how to style them appropriately. Buying an entirely new wardrobe just for an internship didn't feel like the right solution either.&lt;/p&gt;

&lt;p&gt;And that gave me an idea.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What if she could create more outfits from the clothes she already owns?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;That's how &lt;strong&gt;ClosetBestie&lt;/strong&gt; came to life! &lt;/p&gt;

&lt;p&gt;I wanted to build something that feels less like a fashion judge and more like a friend helping you figure out what to wear.&lt;/p&gt;

&lt;h3&gt;
  
  
  What ClosetBestie Does
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt; &lt;strong&gt;Wardrobe Audit:&lt;/strong&gt; Classifies clothes as &lt;em&gt;Ready&lt;/em&gt;, &lt;em&gt;Ready with a fix&lt;/em&gt;, or &lt;em&gt;Weekend only&lt;/em&gt; based on the user's office rules.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Smart Styling:&lt;/strong&gt; Suggests practical fixes, like layering a sleeveless top with a blazer when it meets the dress code.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Weekly Outfit Planner:&lt;/strong&gt; Creates Monday-to-Friday outfit plans with variety and fewer repeated pieces.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Smart Gap Finder:&lt;/strong&gt; Identifies one versatile item that could unlock more outfit combinations instead of suggesting an entire shopping list.&lt;/li&gt;
&lt;li&gt; &lt;strong&gt;Gemma 3 Stylist Notes:&lt;/strong&gt; Adds friendly, encouraging daily notes to make the experience feel more personal.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The goal is simple: more outfit possibilities, fewer unnecessary purchases, and a little more confidence.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why I Built This
&lt;/h2&gt;

&lt;p&gt;Starting somewhere new is exciting, but it can also be overwhelming. You want to make a good impression, feel comfortable, and find your place. I didn't want choosing an outfit to become another thing my friend had to worry about.&lt;/p&gt;

&lt;p&gt;I also didn't want the solution to be, "Just go shopping."&lt;/p&gt;

&lt;p&gt;Sometimes, we don't need more clothes. We just need a little help seeing the possibilities in what we already have.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I built ClosetBestie because I wanted my friend to have one less thing to worry about as she starts this new chapter.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;It's a small project, but it comes from a real problem. And if it helps someone else feel a little more confident getting ready for their day, that makes it even more meaningful to me. &lt;/p&gt;

&lt;h2&gt;
  
  
  Try It Out
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Live App:&lt;/strong&gt; &lt;a href="https://closetbestie-1.onrender.com/" rel="noopener noreferrer"&gt;https://closetbestie-1.onrender.com/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backend API Docs:&lt;/strong&gt; &lt;a href="https://closetbestie.onrender.com/docs" rel="noopener noreferrer"&gt;https://closetbestie.onrender.com/docs&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/kritikamandale/ClosetBestie" rel="noopener noreferrer"&gt;https://github.com/kritikamandale/ClosetBestie&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  A Little About the Tech
&lt;/h2&gt;

&lt;p&gt;ClosetBestie combines &lt;strong&gt;Next.js, FastAPI, Python, and SQLite&lt;/strong&gt; with two AI components.&lt;/p&gt;

&lt;p&gt;I fine-tuned the open-weight Nemotron model using LoRA and Tinker to understand clothing descriptions. I also use Gemma 3 to generate supportive stylist notes.&lt;/p&gt;

&lt;p&gt;In my evaluation on 54 held-out examples, the fine-tuned model improved JSON validity from 13.0% to 51.9% and overall tag accuracy from 9.5% to 47.3% compared with the prompted base model. The results also showed that there's still room for improvement, so I included validation and a manual fallback.&lt;/p&gt;

&lt;p&gt;The outfit planner and gap finder use regular Python logic to keep the important rules consistent.&lt;/p&gt;

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

&lt;p&gt;Open-weight models gave me the opportunity to experiment, fine-tune for a specific task, and evaluate the results instead of relying entirely on a general-purpose model.&lt;/p&gt;

&lt;p&gt;For me, open innovation is about having the freedom to build around a real problem, learn from what works, and be honest about what still needs improvement.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Tinker by Thinking Machines:&lt;/strong&gt; Fine-tuned Nemotron using the Tinker SDK with LoRA and evaluated it against the prompted base model.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Render:&lt;/strong&gt; Deployed the FastAPI backend and Next.js frontend as separate services using a Render Blueprint.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  A Little More Than an Outfit Planner
&lt;/h2&gt;

&lt;p&gt;I know an app can't take away all the nervousness of starting a first internship. But maybe it can make one small part of the experience easier.&lt;/p&gt;

&lt;p&gt;Maybe it can help someone rediscover a favourite outfit, try a combination they hadn't considered, or start the morning feeling a little more prepared.&lt;/p&gt;

&lt;p&gt;That's what I wanted to build for my friend. Not another reason to buy something new, but a little reminder that what you already own can be enough.&lt;/p&gt;

&lt;p&gt;More outfits from the clothes you already own. Fewer unnecessary purchases. And a little more confidence for whatever comes next.&lt;/p&gt;

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