<?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: E. Smith</title>
    <description>The latest articles on DEV Community by E. Smith (@eszter_kovacs_37c655dc5e9).</description>
    <link>https://dev.to/eszter_kovacs_37c655dc5e9</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%2F1837887%2Fe1c694b7-6e8b-42a7-8976-078c022a645f.jpg</url>
      <title>DEV Community: E. Smith</title>
      <link>https://dev.to/eszter_kovacs_37c655dc5e9</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/eszter_kovacs_37c655dc5e9"/>
    <language>en</language>
    <item>
      <title>Haunted Haiku: a local Gemma 3 poet that paints, whispers and hums its own ghost story (Ghost in the Machine)</title>
      <dc:creator>E. Smith</dc:creator>
      <pubDate>Wed, 28 Oct 2026 15:44:37 +0000</pubDate>
      <link>https://dev.to/eszter_kovacs_37c655dc5e9/haunted-haiku-a-local-gemma-3-poet-that-paints-whispers-and-hums-its-own-ghost-story-ghost-in-54lh</link>
      <guid>https://dev.to/eszter_kovacs_37c655dc5e9/haunted-haiku-a-local-gemma-3-poet-that-paints-whispers-and-hums-its-own-ghost-story-ghost-in-54lh</guid>
      <description>&lt;p&gt;&lt;em&gt;This is my submission for the &lt;a href="https://dev.to/challenges/hf26"&gt;Hacktoberfest Open-Source AI Challenge: Week 4&lt;/a&gt; (theme: Ghost in the Machine).&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Haunted Haiku&lt;/strong&gt; is a very small app. You type a word or a mood ("empty house", "candle", "fog"), and a local model writes an eerie haiku about it. Then the app does three more things with it:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Paints it.&lt;/strong&gt; The same model picks two dark, haunting colours and a few spooky emoji, and the app draws the poem as a card: a two-colour gradient, soft bokeh lights and a big, faint emoji scene behind the text. You can save the card as a PNG.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reads it aloud.&lt;/strong&gt; One button sends the poem to ElevenLabs text-to-speech and plays it back.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hums a soundtrack.&lt;/strong&gt; Another button asks ElevenLabs Music for a 20-second instrumental that fits the poem, and loops it under the reading.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;My reading of the theme: the "ghost in the machine" is the voice that whispers a poem back at you from inside your own laptop, on the thin line between the living and the dead (Halloween, Día de los Muertos). The writing is not a cloud service. It is an open-weight model, Gemma 3, running through Ollama on &lt;code&gt;localhost&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why open, and why local
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The poet works with the Wi-Fi off.&lt;/strong&gt; The only call that writes the poem goes to &lt;code&gt;localhost:11434&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What you type stays on the machine.&lt;/strong&gt; A haiku prompt is not sensitive, but the same code works for things that are, like a remembrance for someone you lost.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The model is a one-line swap.&lt;/strong&gt; &lt;code&gt;HAIKU_MODEL=gemma3:12b&lt;/code&gt; (or any other Ollama model) changes the poet without touching the code.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;To be straight about it: the voice and the music are the exception. They come from ElevenLabs, a hosted service, and need an API key and internet. They are optional buttons. The haiku and the picture work fully offline.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;One Python file, &lt;a href="https://dash.plotly.com/" rel="noopener noreferrer"&gt;Dash&lt;/a&gt; with Dash Mantine Components for the UI, and plain &lt;code&gt;requests&lt;/code&gt; for the model call. No SDK for the model.&lt;/p&gt;

&lt;p&gt;The interesting part is that a single request returns both the poem and the art direction, as JSON:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;PROMPT&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Write an eerie, haunted English {form} about: {topic}. Make it spooky: ghosts, whispers, candlelight, fog, &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;the thin line between the living and the dead (Halloween and Dia de los Muertos spirit). &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Also design a picture for it. &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Reply with JSON only: {{&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;lines&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;: [{n} strings], &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;emoji&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;: &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;1 to 3 spooky emoji that show the scene&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;, &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;from&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;: a colour, &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;to&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s"&gt;: a different colour}}. Colours must come from: &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;, &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;COLORS&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;. &lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Pick dark, haunting colours that match the mood (e.g. midnight indigo, deep violet, candle orange).&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;haiku&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;topic&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;n&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;requests&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;post&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;OLLAMA_URL&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt;/api/chat&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;timeout&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;120&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;model&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;MODEL&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;stream&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;False&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;format&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;json&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;options&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;temperature&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mf"&gt;0.9&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;messages&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;role&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;user&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;PROMPT&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;format&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;topic&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;topic&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;form&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;form&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;n&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;n&lt;/span&gt;&lt;span class="p"&gt;)}]})&lt;/span&gt;
    &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;raise_for_status&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
    &lt;span class="n"&gt;d&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;loads&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;json&lt;/span&gt;&lt;span class="p"&gt;()[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;message&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt;
    &lt;span class="bp"&gt;...&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Setting &lt;code&gt;"format": "json"&lt;/code&gt; makes Ollama constrain the output to valid JSON. The colours are limited to Mantine's palette names, and the "picture" is plain CSS. There is no image model involved: a gradient, soft light spots, a large translucent emoji and the poem on top. You can also pick the colour yourself, or leave the switch on and let the model choose. A small stepper changes the poem from 1 to 10 lines.&lt;/p&gt;

&lt;h3&gt;
  
  
  The voice and the music
&lt;/h3&gt;

&lt;p&gt;Both are a few lines with the ElevenLabs Python SDK, and both return an mp3 as a data URI, so the browser can play it with a plain &lt;code&gt;html.Audio&lt;/code&gt; and nothing is written to disk:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;speak&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;ElevenLabs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getenv&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;ELEVENLABS_API_KEY&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
    &lt;span class="n"&gt;audio&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;text_to_speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;convert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;voice_id&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;ELEVEN_VOICE&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;model_id&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;eleven_v4&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;data:audio/mpeg;base64,&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;base64&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;b64encode&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;b&lt;/span&gt;&lt;span class="sh"&gt;""&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;audio&lt;/span&gt;&lt;span class="p"&gt;)).&lt;/span&gt;&lt;span class="nf"&gt;decode&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;music&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;ElevenLabs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getenv&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;ELEVENLABS_API_KEY&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
    &lt;span class="n"&gt;audio&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;music&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;compose&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;prompt&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Eerie, haunted instrumental ambient music, slow and quiet, for this poem: &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;music_length_ms&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;20000&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;force_instrumental&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;data:audio/mpeg;base64,&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;base64&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;b64encode&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;b&lt;/span&gt;&lt;span class="sh"&gt;""&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;audio&lt;/span&gt;&lt;span class="p"&gt;)).&lt;/span&gt;&lt;span class="nf"&gt;decode&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The poem itself is the music prompt, so a poem about fog and a poem about candlelight get different soundtracks. &lt;code&gt;force_instrumental=True&lt;/code&gt; keeps the track from singing over the voice.&lt;/p&gt;

&lt;h2&gt;
  
  
  Things that went wrong (and what I did about them)
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The model does not always follow the schema.&lt;/strong&gt; I asked for &lt;code&gt;"emoji": "1 to 3 emoji"&lt;/code&gt; as a string. Sometimes it returns a &lt;em&gt;list&lt;/em&gt; instead. My first version printed &lt;code&gt;"['🌧️', '🍂', "&lt;/code&gt; into the picture. The fix is three lines: if it is a list, join it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Colours need validating.&lt;/strong&gt; The model can answer with a colour that is not in the palette, or the same colour twice. I check both and fall back to a default pair.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;5-7-5 is a hint, not a guarantee.&lt;/strong&gt; Gemma does not count syllables reliably. Most haiku come out close, but not all are exactly 5-7-5.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;My app said "needs a paid plan" for every music error, and it was wrong.&lt;/strong&gt; Reading aloud worked, but music failed. The real error (401, &lt;code&gt;missing_permissions&lt;/code&gt;) said my ElevenLabs API key had no &lt;code&gt;music_generation&lt;/code&gt; permission. Keys can be restricted per feature, and mine only had text-to-speech. I turned the permission on, and the same call returned audio. If I did it again I would show the API's own message in the app instead of a guess.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Run it
&lt;/h2&gt;

&lt;p&gt;Code: &lt;a href="https://github.com/banana0000/haiku" rel="noopener noreferrer"&gt;banana0000/haiku&lt;/a&gt;&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/banana0000/haiku
&lt;span class="nb"&gt;cd &lt;/span&gt;haiku
ollama pull gemma3:4b
pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; requirements.txt
&lt;span class="c"&gt;# optional, for the voice and the music:&lt;/span&gt;
&lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s2"&gt;"ELEVENLABS_API_KEY=your-key"&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; .env
python app.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then open &lt;code&gt;http://localhost:8063&lt;/code&gt;. There is no login. Type a word, press &lt;strong&gt;Summon a haiku&lt;/strong&gt; (or Enter), then try &lt;strong&gt;Read aloud&lt;/strong&gt;, &lt;strong&gt;Add music&lt;/strong&gt; and &lt;strong&gt;Save as PNG&lt;/strong&gt;. The music button needs an API key with the Music Generation permission.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; Gemma 3 (&lt;code&gt;gemma3:4b&lt;/code&gt;) writes the poem and chooses its colours and emoji. It is the only model that writes anything.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of ElevenLabs:&lt;/strong&gt; the poem is read aloud with ElevenLabs text-to-speech, and ElevenLabs Music composes a short instrumental soundtrack from the poem itself.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Reflection
&lt;/h2&gt;

&lt;p&gt;The pleasant surprise was how far a small model goes when you ask it for &lt;em&gt;structure&lt;/em&gt; instead of just text. One request, one JSON object, and the model is both the poet and the art director. Adding a voice and a soundtrack changed how the whole thing feels: a card on a screen is a picture, but a card that whispers its own poem over a low drone is a little ghost. The unpleasant surprise was how often a small model almost follows a schema. Validating its output is not optional.&lt;/p&gt;

</description>
      <category>hf26challenge</category>
      <category>ollama</category>
      <category>ai</category>
      <category>elevenlabs</category>
    </item>
    <item>
      <title>Receipt Box: keeping freelancers' receipts private with Gemma, built for Hacktoberfest</title>
      <dc:creator>E. Smith</dc:creator>
      <pubDate>Mon, 26 Oct 2026 16:03:24 +0000</pubDate>
      <link>https://dev.to/eszter_kovacs_37c655dc5e9/receipt-box-keeping-freelancers-receipts-private-with-gemma-built-for-hacktoberfest-1aa7</link>
      <guid>https://dev.to/eszter_kovacs_37c655dc5e9/receipt-box-keeping-freelancers-receipts-private-with-gemma-built-for-hacktoberfest-1aa7</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/mlh-hackathon"&gt;MLH x DEV Writing Challenge&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Receipt Box&lt;/strong&gt; is a receipt and invoice sorter for freelancers that runs entirely on your own laptop.&lt;/p&gt;

&lt;p&gt;The idea came from the end-of-quarter ritual every freelancer knows: a folder of crumpled photos, PDF invoices in several currencies, and an accountant asking "which of these have VAT?" Those documents are some of the most sensitive "boring" data you have: who you work for, where you travel, what you pay for. I did not want to ask anyone to upload them to a cloud service, so the whole thing runs locally.&lt;/p&gt;

&lt;p&gt;Drop in photos or PDFs and a vision model reads each one: vendor, date, total, VAT, currency and a category (Software, Travel, Meals...). You get:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a dashboard with total spend, VAT to reclaim, and spend by category and month;&lt;/li&gt;
&lt;li&gt;a VAT-by-quarter table, the number a freelancer actually hands to the accountant;&lt;/li&gt;
&lt;li&gt;an editable table, where changing a vendor's category once teaches Receipt Box to sort that vendor the same way from then on;&lt;/li&gt;
&lt;li&gt;duplicate detection, multi-currency totals, and CSV / Excel export.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The biggest thing I learned: a small model is allowed to be wrong, as long as fixing it is cheap. A 4B model misreads the odd crumpled receipt, so the correction table and the learned vendor rules are what make the tool usable, more than the model itself.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Live app:&lt;/strong&gt; &lt;a href="https://receipt-box-69v1.onrender.com" rel="noopener noreferrer"&gt;https://receipt-box-69v1.onrender.com&lt;/a&gt; (click &lt;strong&gt;Load demo data&lt;/strong&gt; to explore the dashboard)&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%2Fv0deq3v4q4bf902mzps4.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%2Fv0deq3v4q4bf902mzps4.png" alt="Receipt Box dashboard" width="800" height="400"&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%2Flb40gfqsm42nnn1vjkwx.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%2Flb40gfqsm42nnn1vjkwx.png" alt="Receipt Box receipts table" width="800" height="564"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The repo includes three fictional sample documents in &lt;code&gt;samples/&lt;/code&gt; (a PDF invoice and two receipt images in different currencies). Drop them in and the dashboard fills in as each one is read.&lt;/p&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/banana0000" rel="noopener noreferrer"&gt;
        banana0000
      &lt;/a&gt; / &lt;a href="https://github.com/banana0000/receipt-box" rel="noopener noreferrer"&gt;
        receipt-box
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Private receipt &amp;amp; invoice sorter for freelancers: Gemma 3 reads your receipts locally via Ollama. Editable table, learned categories, VAT by quarter, multi-currency, CSV/Excel export.
    &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;Receipt Box&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;A receipt and invoice sorter for freelancers that runs entirely on your laptop
Photos and PDFs go to an open-weight vision model (Gemma 3 via Ollama), and the result is a
categorised expense dashboard plus a CSV or Excel file for your accountant
Nothing is uploaded anywhere.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Features&lt;/h2&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;Reads receipt photos (JPG / PNG) and PDF invoices (text layer or scanned) with Gemma 3.&lt;/li&gt;
&lt;li&gt;Dashboard: total spend, VAT to reclaim, spend by category and by month, VAT by quarter.&lt;/li&gt;
&lt;li&gt;Editable table: fix any cell, delete rows. Changing a category teaches the app: that vendor is
sorted the same way from now on (saved in &lt;code&gt;rules.json&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;Possible duplicates are highlighted.&lt;/li&gt;
&lt;li&gt;Several currencies: pick one, or "All currencies" to see them side by side. Amounts are never converted.&lt;/li&gt;
&lt;li&gt;Export to CSV or Excel (receipts, spend per category, VAT per quarter).&lt;/li&gt;
&lt;li&gt;Optional line-item extraction (slower).&lt;/li&gt;
&lt;li&gt;Files are read one at…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/banana0000/receipt-box" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;h2&gt;
  
  
  Partner Technologies
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Gemma 3 (Google DeepMind), through Ollama.&lt;/strong&gt; The entire product is built on Gemma 3 4B with vision, running on the user's machine. The only network call the app makes goes to &lt;code&gt;localhost&lt;/code&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Each image is sent to Gemma with a JSON schema (&lt;code&gt;vendor&lt;/code&gt;, &lt;code&gt;date&lt;/code&gt;, &lt;code&gt;total&lt;/code&gt;, &lt;code&gt;vat&lt;/code&gt;, &lt;code&gt;currency&lt;/code&gt;, &lt;code&gt;category&lt;/code&gt;), and the category is constrained to a fixed list. The reply always parses, so there is no regex and no retry loop.&lt;/li&gt;
&lt;li&gt;PDFs take two paths: if there is a text layer, the text goes straight to the model, which is faster; scanned PDFs are rendered to an image with PyMuPDF and go through the vision path.&lt;/li&gt;
&lt;li&gt;On a CPU a receipt takes roughly 15-20 seconds, so files are processed in a queue and the page shows "Reading 2 of 5: name.png" instead of freezing.&lt;/li&gt;
&lt;li&gt;The prompt tells the model to use only what is printed and never guess. A bigger Gemma or a GPU is a one-line change (&lt;code&gt;RECEIPT_MODEL&lt;/code&gt;).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Render.&lt;/strong&gt; The app also deploys from a single Blueprint (&lt;code&gt;render.yaml&lt;/code&gt;): a private Ollama service with Gemma 3 baked into its Docker image, and the web app talking to it over Render's private network, so the model is never exposed to the internet. Gemma 3 4B needs about 4-5 GB of RAM, so the Ollama service needs a large instance, which was the main thing I had to plan around.&lt;/p&gt;

&lt;p&gt;The rest is a Dash app with Plotly charts, pandas for the numbers, openpyxl for the Excel export, and pytest tests for the data layer (duplicates, quarters, rules, exports).&lt;/p&gt;

&lt;h2&gt;
  
  
  Hackathon Experience
&lt;/h2&gt;

&lt;p&gt;I built Receipt Box for the &lt;strong&gt;Hacktoberfest Open-Source AI Challenge, Week 2: Side Hustle&lt;/strong&gt; (October 12 to 18, 2026), the online challenge run by DEV together with MLH. The prompt was to build something with open-weight models that someone would actually pay for, and an invoice and receipt sorter that runs locally was one of the ideas in it, which pushed me to take the privacy question seriously instead of treating it as a side note.&lt;/p&gt;

&lt;p&gt;Having a one-week window with a clear theme kept the scope honest: I cut features that did not help the "hand this to my accountant" moment, and spent the time on the correction table, quarterly VAT and the progress indicator.&lt;/p&gt;

</description>
      <category>mlhacks</category>
      <category>gemma</category>
      <category>python</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Slothstep: a lazy-friendly app that turns "ugh, I'll do it later" into three tiny steps with gemma</title>
      <dc:creator>E. Smith</dc:creator>
      <pubDate>Fri, 23 Oct 2026 14:43:22 +0000</pubDate>
      <link>https://dev.to/eszter_kovacs_37c655dc5e9/slothstep-a-lazy-friendly-app-that-turns-ugh-ill-do-it-later-into-three-tiny-steps-5cd3</link>
      <guid>https://dev.to/eszter_kovacs_37c655dc5e9/slothstep-a-lazy-friendly-app-that-turns-ugh-ill-do-it-later-into-three-tiny-steps-5cd3</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest"&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;Slothstep&lt;/strong&gt; is a tiny web app for the thing sloths do best: not panicking.&lt;/p&gt;

&lt;p&gt;You type whatever you keep putting off ("clean the whole flat before Saturday", "write that difficult email to my boss", "sort the files on my laptop") and an open-weight model answers with exactly &lt;strong&gt;three steps&lt;/strong&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;one you can start &lt;em&gt;right now&lt;/em&gt; and finish in two minutes or less,&lt;/li&gt;
&lt;li&gt;one that follows naturally,&lt;/li&gt;
&lt;li&gt;one that finishes the job.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Every step has a realistic time estimate, and you can tick them off as you go. The point is not a perfect plan. The point is that the first step is so small there is no excuse left not to start.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The effort it saves:&lt;/strong&gt; the hardest part of a dreaded task is rarely the task, it is the staring. Slothstep replaces "where do I even begin?" with one concrete, two-minute action. (On my machine a plan takes about 25 seconds to generate with a 4B model.)&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;Live app (demo mode with sample plans, because a free host cannot run a model): &lt;a href="https://slothstep.onrender.com/" rel="noopener noreferrer"&gt;https://slothstep.onrender.com/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Real thing: clone the repo, &lt;code&gt;ollama pull gemma3:4b&lt;/code&gt;, &lt;code&gt;pip install -r requirements.txt&lt;/code&gt;, &lt;code&gt;python app.py&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Code: &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/banana0000" rel="noopener noreferrer"&gt;
        banana0000
      &lt;/a&gt; / &lt;a href="https://github.com/banana0000/slothstep" rel="noopener noreferrer"&gt;
        slothstep
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Slothstep: turn a task you keep putting off into three tiny steps with an open-weight model (Gemma 3 + Ollama). Dash app.
    &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;🦥 Slothstep&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;Describe a task you keep putting off and get &lt;strong&gt;three tiny, realistic steps&lt;/strong&gt;. The first one takes two minutes or less, so there is no excuse left not to start.&lt;/p&gt;
&lt;p&gt;Built for the Hacktoberfest Open-Source AI Challenge (Week 3: &lt;em&gt;Slow and Steady&lt;/em&gt;) with an open-weight model (&lt;a href="https://ai.google.dev/gemma" rel="nofollow noopener noreferrer"&gt;Gemma 3&lt;/a&gt;) served by &lt;a href="https://ollama.com/" rel="nofollow noopener noreferrer"&gt;Ollama&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Live demo:&lt;/strong&gt; &lt;a href="https://slothstep.onrender.com" rel="nofollow noopener noreferrer"&gt;https://slothstep.onrender.com&lt;/a&gt;&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Features&lt;/h2&gt;
&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;Three steps with time estimates, validated so the model cannot return malformed plans&lt;/li&gt;
&lt;li&gt;Tick steps off as you go&lt;/li&gt;
&lt;li&gt;Light and dark mode (remembered in the browser)&lt;/li&gt;
&lt;li&gt;A badge showing whether the model server is running&lt;/li&gt;
&lt;li&gt;Runs fully offline with a local model, or against any OpenAI-compatible host serving an open-weight model&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Stack: &lt;a href="https://dash.plotly.com/" rel="nofollow noopener noreferrer"&gt;Dash&lt;/a&gt; + &lt;a href="https://www.dash-mantine-components.com/" rel="nofollow noopener noreferrer"&gt;Dash Mantine Components&lt;/a&gt;, Python.&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Run locally&lt;/h2&gt;
&lt;/div&gt;

&lt;div class="snippet-clipboard-content notranslate position-relative overflow-auto"&gt;&lt;pre class="notranslate"&gt;&lt;code&gt;ollama pull gemma3:4b
pip install -r requirements.txt
python app.py
&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;
&lt;p&gt;Open &lt;a href="http://127.0.0.1:8050" rel="nofollow noopener noreferrer"&gt;http://127.0.0.1:8050&lt;/a&gt;.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Configuration&lt;/h2&gt;

&lt;/div&gt;

&lt;p&gt;&lt;/p&gt;&lt;div class="table-wrapper-paragraph"&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;table&gt;

&lt;thead&gt;

&lt;tr&gt;

&lt;th&gt;Variable&lt;/th&gt;

&lt;th&gt;Default&lt;/th&gt;

&lt;th&gt;Purpose&lt;/th&gt;

&lt;/tr&gt;

&lt;/thead&gt;

&lt;tbody&gt;

&lt;tr&gt;

&lt;td&gt;&lt;code&gt;OLLAMA_MODEL&lt;/code&gt;&lt;/td&gt;

&lt;td&gt;&lt;code&gt;gemma3:4b&lt;/code&gt;&lt;/td&gt;

&lt;td&gt;Model name (Ollama tag, or model id on&lt;/td&gt;

&lt;/tr&gt;

&lt;/tbody&gt;

&lt;/table&gt;&lt;/div&gt;…&lt;p&gt;&lt;/p&gt;&lt;/div&gt;
&lt;br&gt;
  &lt;/div&gt;
&lt;br&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/banana0000/slothstep" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;br&gt;
&lt;/div&gt;

&lt;/li&gt;

&lt;/ul&gt;


&lt;div class="crayons-card c-embed text-styles text-styles--secondary"&gt;
    &lt;div class="c-embed__content"&gt;
      &lt;div class="c-embed__body flex items-center justify-between"&gt;
        &lt;a href="https://slothstep.onrender.com/" rel="noopener noreferrer" class="c-link fw-bold flex items-center"&gt;
          &lt;span class="mr-2"&gt;slothstep.onrender.com&lt;/span&gt;
          

        &lt;/a&gt;
      &lt;/div&gt;
    &lt;/div&gt;
&lt;/div&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%2F957pelw3jd792502abvw.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%2F957pelw3jd792502abvw.png" alt=" " width="800" height="372"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The free host may take about a minute to wake up on the first visit.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;UI:&lt;/strong&gt; &lt;a href="https://dash.plotly.com/" rel="noopener noreferrer"&gt;Dash&lt;/a&gt; with &lt;a href="https://www.dash-mantine-components.com/" rel="noopener noreferrer"&gt;Dash Mantine Components&lt;/a&gt; (light and dark mode), a single Python file.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; &lt;a href="https://ai.google.dev/gemma" rel="noopener noreferrer"&gt;Gemma 3&lt;/a&gt; (4B) running locally through &lt;a href="https://ollama.com/" rel="noopener noreferrer"&gt;Ollama&lt;/a&gt;. The same code can also talk to any OpenAI-compatible host serving an open-weight model. The free hosted demo runs in a clearly labelled demo mode with sample plans, since a free host cannot run a model.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reliable output:&lt;/strong&gt; the model is asked for JSON only, then the app validates it (exactly three steps, non-empty text, positive integer minutes) and shows a friendly error instead of garbage if the model misbehaves.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Prompting:&lt;/strong&gt; low temperature, a hard limit of two minutes for step one, and an explicit "do not invent missing details" instruction so it doesn't hallucinate a kitchen you don't have.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deploy:&lt;/strong&gt; a &lt;code&gt;render.yaml&lt;/code&gt; Blueprint for a free Render web service running with &lt;code&gt;DEMO_MODE=true&lt;/code&gt;. A real model needs a private Ollama service on a paid instance, which the repo's &lt;code&gt;ollama/Dockerfile&lt;/code&gt; supports.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;What people procrastinate on is personal: unanswered emails, messy folders, awkward conversations. I don't want those sent to a closed API.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Private by default:&lt;/strong&gt; run it locally and your to-dos never leave your machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Works offline:&lt;/strong&gt; no internet, no problem. Ollama and a 4B model are enough.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Set up once, free forever:&lt;/strong&gt; no subscription, no per-token bill for something you use ten times a week.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Swappable:&lt;/strong&gt; because Slothstep speaks both Ollama and the OpenAI-compatible API, moving from a local Gemma to a hosted open-weight model is two environment variables, not a rewrite.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A closed model might write slightly nicer sentences. But for a three-line plan, a small open model is more than good enough, and I keep control of my data.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; Gemma 3 is the model at the core of the app.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render:&lt;/strong&gt; the app is deployed from a &lt;code&gt;render.yaml&lt;/code&gt; Blueprint.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Thanks for reading, and please take it slow. 🦥&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>hacktoberfest</category>
      <category>gemma</category>
    </item>
    <item>
      <title>Receipt Box: a private receipt sorter for freelancers, powered by Gemma</title>
      <dc:creator>E. Smith</dc:creator>
      <pubDate>Sat, 17 Oct 2026 07:43:55 +0000</pubDate>
      <link>https://dev.to/eszter_kovacs_37c655dc5e9/receipt-box-a-private-receipt-sorter-for-freelancers-powered-by-gemma-136m</link>
      <guid>https://dev.to/eszter_kovacs_37c655dc5e9/receipt-box-a-private-receipt-sorter-for-freelancers-powered-by-gemma-136m</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;&lt;strong&gt;Receipt Box&lt;/strong&gt; is a receipt and invoice sorter for freelancers that runs entirely on your own laptop.&lt;/p&gt;

&lt;p&gt;Every freelancer knows the end-of-quarter ritual: a folder of crumpled photos, PDF invoices in six currencies, and an accountant asking "which of these have VAT?" Receipt Box turns that folder into a clean expense dashboard in a few minutes. Drop in photos or PDFs and a vision model reads each one: vendor, date, total, VAT, currency, and a category such as Software, Travel or Meals. You get:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;A dashboard&lt;/strong&gt;: total spend, VAT to reclaim, spend by category and by month.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A VAT-by-quarter table&lt;/strong&gt;, because that is the number a freelancer actually has to hand over.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;An editable table.&lt;/strong&gt; A 4B model will sometimes be wrong, so every cell is fixable. Change a vendor's category once and Receipt Box remembers it: that vendor is sorted the same way from then on.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Duplicate detection&lt;/strong&gt;, so the same invoice uploaded twice gets flagged instead of counted twice.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multi-currency handling&lt;/strong&gt;, with a currency switch instead of silently adding euros to forints.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Exports&lt;/strong&gt; to CSV and to an Excel workbook (receipts, spend per category, VAT per quarter) that goes straight to the accountant.&lt;/li&gt;
&lt;li&gt;Optional &lt;strong&gt;line-item extraction&lt;/strong&gt; for when you want to see what was bought, not just the total.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Who would pay?&lt;/strong&gt; Freelancers and tiny agencies who want bookkeeping help but cannot (or should not) upload client invoices, travel documents and personal receipts to a cloud service. The same goes for the accountants who serve them: with many clients' paperwork on their desk, "nothing leaves this machine" is a feature they can promise. I see it as a one-time licence for freelancers, or a small monthly fee per accountant seat.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Live app:&lt;/strong&gt; &lt;a href="https://receipt-box-69v1.onrender.com" rel="noopener noreferrer"&gt;https://receipt-box-69v1.onrender.com&lt;/a&gt; (hosted on Render; click &lt;strong&gt;Load demo data&lt;/strong&gt; to explore the dashboard)&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%2Fv0deq3v4q4bf902mzps4.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%2Fv0deq3v4q4bf902mzps4.png" alt=" " width="800" height="400"&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%2Flb40gfqsm42nnn1vjkwx.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%2Flb40gfqsm42nnn1vjkwx.png" alt=" " width="800" height="564"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The repo includes three fictional sample documents in &lt;code&gt;samples/&lt;/code&gt; (a PDF invoice and two receipt images in different currencies). Drop them in and the dashboard fills in as each document is read.&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/banana0000" rel="noopener noreferrer"&gt;
        banana0000
      &lt;/a&gt; / &lt;a href="https://github.com/banana0000/receipt-box" rel="noopener noreferrer"&gt;
        receipt-box
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Private receipt &amp;amp; invoice sorter for freelancers: Gemma 3 reads your receipts locally via Ollama. Editable table, learned categories, VAT by quarter, multi-currency, CSV/Excel export.
    &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;Receipt Box&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;A receipt and invoice sorter for freelancers that runs entirely on your laptop
Photos and PDFs go to an open-weight vision model (Gemma 3 via Ollama), and the result is a
categorised expense dashboard plus a CSV or Excel file for your accountant
Nothing is uploaded anywhere.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Features&lt;/h2&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;Reads receipt photos (JPG / PNG) and PDF invoices (text layer or scanned) with Gemma 3.&lt;/li&gt;
&lt;li&gt;Dashboard: total spend, VAT to reclaim, spend by category and by month, VAT by quarter.&lt;/li&gt;
&lt;li&gt;Editable table: fix any cell, delete rows. Changing a category teaches the app: that vendor is
sorted the same way from now on (saved in &lt;code&gt;rules.json&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;Possible duplicates are highlighted.&lt;/li&gt;
&lt;li&gt;Several currencies: pick one, or "All currencies" to see them side by side. Amounts are never converted.&lt;/li&gt;
&lt;li&gt;Export to CSV or Excel (receipts, spend per category, VAT per quarter).&lt;/li&gt;
&lt;li&gt;Optional line-item extraction (slower).&lt;/li&gt;
&lt;li&gt;Files are read one at…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/banana0000/receipt-box" 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;The whole project is built around one open-weight model: &lt;strong&gt;Gemma 3 (4B, vision)&lt;/strong&gt;, running locally through &lt;strong&gt;Ollama&lt;/strong&gt;. The only network call the app makes goes to &lt;code&gt;localhost&lt;/code&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Structured output.&lt;/strong&gt; Each image is sent to Gemma with a JSON schema (&lt;code&gt;vendor&lt;/code&gt;, &lt;code&gt;date&lt;/code&gt;, &lt;code&gt;total&lt;/code&gt;, &lt;code&gt;vat&lt;/code&gt;, &lt;code&gt;currency&lt;/code&gt;, &lt;code&gt;category&lt;/code&gt;), and the category is constrained to a fixed list. The reply always parses, with no regex and no retry loop.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Two paths for PDFs.&lt;/strong&gt; If a PDF has a text layer, the text goes to the model directly, which is faster. Scanned PDFs are rendered to an image (PyMuPDF) and go through the vision path.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Queue with visible progress.&lt;/strong&gt; On a CPU a receipt takes roughly 15-20 seconds, so files are processed one at a time and the page shows "Reading 2 of 5: name.png" instead of freezing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The model is allowed to be wrong.&lt;/strong&gt; The prompt tells it to use only what is printed and not to guess. After that, the human-in-the-loop table and the learned vendor rules do the rest. In my tests the three sample documents were read correctly, and for real-world crumpled receipts the correction table is what makes the tool usable.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dashboard.&lt;/strong&gt; A Dash app with Plotly charts and pandas for the numbers (quarterly VAT, duplicates, per-currency totals). Excel export uses openpyxl.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tests.&lt;/strong&gt; The data layer (duplicates, quarters, rules, exports) is covered with pytest.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Hosting.&lt;/strong&gt; The same app also deploys to Render from a Blueprint (&lt;code&gt;render.yaml&lt;/code&gt;): a private Ollama service with Gemma 3 baked into its Docker image, and the web app talking to it over Render's private network, so the model is never exposed to the internet. Gemma 3 4B needs about 4-5 GB of RAM, so the Ollama service runs on a large instance. On a laptop none of this is needed.&lt;/p&gt;

&lt;p&gt;Honest limits: a 4B model on a CPU is slow, and line-item extraction is slower still, so it is off by default. A bigger Gemma or a GPU is a one-line change (&lt;code&gt;RECEIPT_MODEL&lt;/code&gt;).&lt;/p&gt;

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

&lt;p&gt;Receipts are about the most sensitive "boring" data a freelancer has: who you work for, where you travel, what you pay for. A closed API would force a trade-off between convenience and privacy, and that is exactly where I would lose the customers who care about it most.&lt;/p&gt;

&lt;p&gt;An open-weight model removes the trade-off. Because Gemma runs on the user's own machine:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the documents never leave the laptop, so there is nothing to leak, retain or train on;&lt;/li&gt;
&lt;li&gt;there is no per-receipt API bill, so the price of the product is not tied to someone else's pricing;&lt;/li&gt;
&lt;li&gt;it keeps working offline, on a train or in a client's office with no Wi-Fi;&lt;/li&gt;
&lt;li&gt;the user can swap the model by changing one environment variable as better open models arrive.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A small open model that runs on a laptop is what makes "private by design" a real promise here, not a marketing line.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Gemma&lt;/strong&gt;: the entire product is built on Gemma 3 running locally.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Render&lt;/strong&gt;: the web app and a private Gemma/Ollama service are deployed together from one Blueprint.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>mlhacks</category>
    </item>
    <item>
      <title>Exam AI – something I built during the MeDo hackathon</title>
      <dc:creator>E. Smith</dc:creator>
      <pubDate>Mon, 04 May 2026 07:30:20 +0000</pubDate>
      <link>https://dev.to/eszter_kovacs_37c655dc5e9/exam-ai-something-i-built-during-the-medo-hackathon-23a1</link>
      <guid>https://dev.to/eszter_kovacs_37c655dc5e9/exam-ai-something-i-built-during-the-medo-hackathon-23a1</guid>
      <description>&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.amazonaws.com%2Fuploads%2Farticles%2F7k0eq811li30ii364je2.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.amazonaws.com%2Fuploads%2Farticles%2F7k0eq811li30ii364je2.png" alt=" " width="800" height="374"&gt;&lt;/a&gt;&lt;br&gt;
🚀 During the MeDo hackathon,I built something surprisingly simple — but genuinely useful: &lt;strong&gt;Exam AI&lt;/strong&gt; 📚🤖&lt;/p&gt;

&lt;p&gt;The idea came from a pretty common frustration.&lt;/p&gt;

&lt;p&gt;A lot of studying turns into passive repetition:&lt;br&gt;
reading the same notes again and again, scrolling slides, highlighting things… and hoping it somehow sticks.&lt;/p&gt;

&lt;p&gt;I wanted to build something more active.&lt;br&gt;
Something that actually makes you &lt;em&gt;think&lt;/em&gt; while learning.&lt;/p&gt;

&lt;p&gt;So I created Exam AI. This AI tool is very easy to use.&lt;/p&gt;

&lt;p&gt;You give it a topic, and it:&lt;br&gt;
✨ generates exam-style questions&lt;br&gt;
✨ lets you answer interactively&lt;br&gt;
✨ explains the answers clearly&lt;br&gt;
✨ helps you dig deeper when something doesn’t fully click&lt;/p&gt;

&lt;p&gt;It feels less like a quiz app and more like studying with someone who keeps asking the right questions.&lt;/p&gt;

&lt;h2&gt;
  
  
  ⚙️ Building it
&lt;/h2&gt;

&lt;p&gt;I used MeDo to build the project, which let me move really fast.&lt;/p&gt;

&lt;p&gt;Instead of spending the whole hackathon dealing with setup or infrastructure, I could focus on:&lt;br&gt;
🧠 refining prompts&lt;br&gt;
🧠 improving explanations&lt;br&gt;
🧠 making the experience actually helpful&lt;/p&gt;

&lt;p&gt;Most of the work wasn’t technical complexity — it was figuring out how to make the AI respond in a way that feels useful and natural.&lt;/p&gt;

&lt;h2&gt;
  
  
  🧩 The difficult part
&lt;/h2&gt;

&lt;p&gt;The biggest challenge was keeping things consistent.&lt;/p&gt;

&lt;p&gt;With AI, small changes in prompts can completely change the quality of the output.&lt;br&gt;
Some questions were great. Others… not so much 😅&lt;/p&gt;

&lt;p&gt;We also had to constantly fight the urge to add more and more features.&lt;/p&gt;

&lt;p&gt;In the end, keeping the product focused was probably one of the best decisions we made.&lt;/p&gt;

&lt;h2&gt;
  
  
  📚 Biggest takeaway
&lt;/h2&gt;

&lt;p&gt;Hackathons really remind you that:&lt;br&gt;
⚡ speed matters&lt;br&gt;
⚡ simple ideas win&lt;br&gt;
⚡ shipping something beats endlessly planning it&lt;/p&gt;

&lt;p&gt;And honestly, AI becomes much more powerful when you focus less on “cool features” and more on helping people in a practical way.&lt;/p&gt;

&lt;p&gt;I am excited to keep improving it — especially around personalization, progress tracking, and adapting to weak areas over time.&lt;/p&gt;

&lt;p&gt;But for now, we’re just happy we turned an idea into something real 🙌&lt;/p&gt;

&lt;p&gt;You can check my app here:&lt;br&gt;
&lt;a href="https://medo.dev/projects/app-bdybwhukvg8x" rel="noopener noreferrer"&gt;https://medo.dev/projects/app-bdybwhukvg8x&lt;/a&gt;&lt;br&gt;
&lt;a href="https://www.youtube.com/watch?v=aFDX8r6-ULY" rel="noopener noreferrer"&gt;https://www.youtube.com/watch?v=aFDX8r6-ULY&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Register here and you can get 300 credit:  pagea.uk/ono2bg&lt;/p&gt;

</description>
      <category>ai</category>
      <category>learning</category>
      <category>builtwithmedo</category>
    </item>
    <item>
      <title>🎃 My Kiroween Hackathon Experience — I Built a Spooky Landing Page with MCP Integration using Kiro IDE</title>
      <dc:creator>E. Smith</dc:creator>
      <pubDate>Wed, 26 Nov 2025 10:17:39 +0000</pubDate>
      <link>https://dev.to/eszter_kovacs_37c655dc5e9/my-kiroween-hackathon-experience-i-built-a-spooky-landing-page-with-mcp-integration-using-kiro-5al1</link>
      <guid>https://dev.to/eszter_kovacs_37c655dc5e9/my-kiroween-hackathon-experience-i-built-a-spooky-landing-page-with-mcp-integration-using-kiro-5al1</guid>
      <description>&lt;p&gt;I joined the Kiroween hackathon with one goal: &lt;strong&gt;build something fun&lt;/strong&gt;.&lt;br&gt;
A few late-night ideas later, I ended up creating a &lt;strong&gt;Halloween-themed, API-powered landing page&lt;/strong&gt; — entirely by myself, inside &lt;strong&gt;Kiro IDE&lt;/strong&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.amazonaws.com%2Fuploads%2Farticles%2F9x3ejlvwbv0jsbix5svg.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.amazonaws.com%2Fuploads%2Farticles%2F9x3ejlvwbv0jsbix5svg.png" alt=" " width="799" height="356"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  👻 What I made
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Spooky Landing Page with MCP Integration&lt;/strong&gt; — an animated, glowing UI that pulls real-time spooky quotes via MCP.&lt;br&gt;
It reacts, flickers, whispers Halloween mood into the screen.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Features:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🔮 MCP + ZenQuotes API connection&lt;/li&gt;
&lt;li&gt;💀 Animated UI states (ghost / moon / crystal ball)&lt;/li&gt;
&lt;li&gt;🎨 DiceBear avatar generation&lt;/li&gt;
&lt;li&gt;⚡ Smooth transitions + spooky sound cues&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Short project — big vibes.&lt;/p&gt;

&lt;h2&gt;
  
  
  🧰 Built in &lt;strong&gt;Kiro IDE&lt;/strong&gt;
&lt;/h2&gt;

&lt;p&gt;I didn’t code this traditionally — I &lt;strong&gt;built it through conversation&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Using Kiro IDE, I generated components, refined effects, and integrated MCP endpoints quickly.&lt;br&gt;
It felt less like typing code, more like &lt;em&gt;designing an idea in real time&lt;/em&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  🏁 Why this mattered to me
&lt;/h2&gt;

&lt;p&gt;I learned how fast a solo developer can move when the tooling flows with creativity.&lt;br&gt;
Kiro IDE helped me think, build, and refine — not just code.&lt;/p&gt;

&lt;p&gt;Small project, but it left a big smile.👻🔥&lt;/p&gt;

</description>
      <category>kiro</category>
      <category>mcp</category>
      <category>vibecoding</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
