<?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: Srilal Sachintha</title>
    <description>The latest articles on DEV Community by Srilal Sachintha (@srilals).</description>
    <link>https://dev.to/srilals</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%2F240793%2Fce277714-26dd-40ce-b3f0-49fb1fe9278b.jpg</url>
      <title>DEV Community: Srilal Sachintha</title>
      <link>https://dev.to/srilals</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/srilals"/>
    <language>en</language>
    <item>
      <title>ghost.py: My Terminal Is Haunted (and It Remembers Me)</title>
      <dc:creator>Srilal Sachintha</dc:creator>
      <pubDate>Sat, 31 Oct 2026 09:10:02 +0000</pubDate>
      <link>https://dev.to/srilals/ghostpy-my-terminal-is-haunted-and-it-remembers-me-4hi8</link>
      <guid>https://dev.to/srilals/ghostpy-my-terminal-is-haunted-and-it-remembers-me-4hi8</guid>
      <description>&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;ghost.py&lt;/code&gt; possesses your shell. You run commands through it as usual, and after each one a ghost whispers back&lt;br&gt;
about what you just did. It's powered by Gemma 3 1B, an open-weight model running locally through Ollama.&lt;/p&gt;

&lt;p&gt;Pick who haunts you:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Ghost&lt;/th&gt;
&lt;th&gt;Personality&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;💀 The Sysadmin of 1999&lt;/td&gt;
&lt;td&gt;Died in the Y2K panic. Tired, judgmental, misses Slackware.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;🌼 La Catrina&lt;/td&gt;
&lt;td&gt;The elegant skeleton of Dia de los Muertos. Treats every command as an offering on an ofrenda.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;📟 The Spirit of COBOL&lt;/td&gt;
&lt;td&gt;Summoned from a 1959 mainframe. SPEAKS ONLY IN UPPERCASE. Reminds you banks still run on it.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;And the haunting goes deeper than one-liners:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🧠 &lt;strong&gt;It remembers you.&lt;/strong&gt; Close the terminal, come back tomorrow, and the ghost greets you by recalling commands
from your last visit.&lt;/li&gt;
&lt;li&gt;😠 &lt;strong&gt;It gets angry.&lt;/strong&gt; Every failed command raises the haunting intensity, from calm to uneasy to furious.
At full fury the screen flickers.&lt;/li&gt;
&lt;li&gt;🕯️ &lt;strong&gt;Seances.&lt;/strong&gt; Type &lt;code&gt;seance &amp;lt;question&amp;gt;&lt;/code&gt; to ask the ghost anything.&lt;/li&gt;
&lt;li&gt;🪦 &lt;strong&gt;A graveyard.&lt;/strong&gt; When you leave, the ghost carves a two-line epitaph for your session into
&lt;code&gt;~/.ghost/graveyard.md&lt;/code&gt;. Over time it becomes a little cemetery of everything you've done in that terminal,
which felt right for Dia de los Muertos: remembering what came before.&lt;/li&gt;
&lt;/ul&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%2Faimvgeyh45vz9rpxnxdb.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%2Faimvgeyh45vz9rpxnxdb.png" alt=" " width="800" height="702"&gt;&lt;/a&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;haunted:WK4&amp;gt; notacommand
'notacommand' is not recognized as an internal or external command
👻 Such a curious deviation. It seems you're attempting to impose order upon the chaos of the departed.

haunted:WK4&amp;gt; exit
👻 "Here lingers bone, a silent plea,
   Remembered echoes, eternally free."
Epitaph carved into ~/.ghost/graveyard.md
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;&amp;gt; python ghost.py --ghost cobol dir /b
👻 A SIMPLE QUERY, STILL, STILL A SIMPLE QUERY.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Code: &lt;a href="https://github.com/SrilalS/ghost-py" rel="noopener noreferrer"&gt;https://github.com/SrilalS/ghost-py&lt;/a&gt;&lt;/p&gt;

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

&lt;p&gt;One Python file, standard library only, talking to Ollama's local &lt;code&gt;/api/chat&lt;/code&gt; endpoint.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Run the command&lt;/strong&gt; with &lt;code&gt;subprocess&lt;/code&gt; and print its output (&lt;code&gt;cd&lt;/code&gt; is handled by the shell itself so it sticks).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Build the ghost's prompt:&lt;/strong&gt; persona + current mood (based on failure count) + memories from past sessions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Send it with recent conversation history&lt;/strong&gt;, so the ghost can build on its earlier comments.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stream the reply&lt;/strong&gt; token by token, printed with a small typewriter delay.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;On exit&lt;/strong&gt;, ask for an epitaph, append it to the graveyard, and save the last few commands to
&lt;code&gt;~/.ghost/memory.json&lt;/code&gt; for next time.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The fun part was how much personality a 1B model has when you give it a strong character and a mood.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Your commands stay on your machine.&lt;/strong&gt; Shell history and output are full of paths, hostnames and secrets,
and this ghost also keeps a memory of them. Sending all of that to a hosted API would be the actually scary
part. Here nothing leaves localhost, and the memory is a plain JSON file you can read or delete.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Works with the lights off.&lt;/strong&gt; No internet, no API key, no bill. It runs on an ordinary laptop.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Swap the ghost.&lt;/strong&gt; One environment variable (&lt;code&gt;GHOST_MODEL&lt;/code&gt;) trades Gemma for any other open model, and each
personality is just a prompt you can edit. A closed service could change or disappear; this ghost stays put.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A 1B model is enough.&lt;/strong&gt; Short, in-character reactions don't need a frontier model, and a tiny local one
answers in about a second.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>hf26challenge</category>
      <category>devchallenge</category>
      <category>opensource</category>
      <category>ai</category>
    </item>
    <item>
      <title>Forkpath: Change One Moment in History, Then Let the AI Do the Homework</title>
      <dc:creator>Srilal Sachintha</dc:creator>
      <pubDate>Tue, 20 Oct 2026 04:08:50 +0000</pubDate>
      <link>https://dev.to/srilals/forkpath-change-one-moment-in-history-then-let-the-ai-do-the-homework-4l47</link>
      <guid>https://dev.to/srilals/forkpath-change-one-moment-in-history-then-let-the-ai-do-the-homework-4l47</guid>
      <description>&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Forkpath&lt;/strong&gt; is an alternate-history explorer. You type one sentence ("What if the Ming treasure fleets kept sailing?"), pick when it happens, and an AI writer maps out what follows, event by event, as a branching tree with the odds at every fork.&lt;/p&gt;

&lt;p&gt;Writing good alternate history is a lot of work. You research the period, think through who wants what, follow second- and third-order effects, and keep half a dozen storylines consistent. Forkpath does that homework for you. You type one line, lean back, and watch the timeline grow.&lt;/p&gt;

&lt;p&gt;What it gives you:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🌳 &lt;strong&gt;A map, not a story.&lt;/strong&gt; Every event is a node. Where outcomes are genuinely uncertain, the writer branches and gives each branch a likelihood. A &lt;strong&gt;Most likely&lt;/strong&gt; view fades every path history probably did not take.&lt;/li&gt;
&lt;li&gt;🧵 &lt;strong&gt;Parallel streams.&lt;/strong&gt; Regions, wars and programs run as separate threads that influence and merge into each other.&lt;/li&gt;
&lt;li&gt;⚖️ &lt;strong&gt;A realism check.&lt;/strong&gt; Language models lean toward tidy, happy endings. The writer checks incentives, capabilities and historical base rates before each event, and an optional second review flags unearned outcomes, anachronisms and numbers out of scale.&lt;/li&gt;
&lt;li&gt;✍️ &lt;strong&gt;You stay the author.&lt;/strong&gt; Edit, regenerate, branch or delete any event. Your edits become canon, and events that depended on them are marked stale.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;Forkpath runs locally, so here's a quick tour in screenshots.&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%2Fvq70qbep195ecdksn5fz.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%2Fvq70qbep195ecdksn5fz.png" alt=" " width="800" height="500"&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%2Fnqtrrt9l4767emzovzcc.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%2Fnqtrrt9l4767emzovzcc.png" alt=" " width="800" height="500"&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%2Fjuznzpe6gnce8h5nhau4.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%2Fjuznzpe6gnce8h5nhau4.png" alt=" " width="800" height="500"&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%2Fgc0dzvb7ym17rmz6jnle.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%2Fgc0dzvb7ym17rmz6jnle.png" alt=" " width="800" height="500"&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%2Fidrq4znccrjwcsa0j1wb.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%2Fidrq4znccrjwcsa0j1wb.png" alt=" " width="800" height="500"&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/SrilalS" rel="noopener noreferrer"&gt;
        SrilalS
      &lt;/a&gt; / &lt;a href="https://github.com/SrilalS/ForkPath" rel="noopener noreferrer"&gt;
        ForkPath
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Change one moment in history and watch the consequences branch out. An AI-powered alternate-history mapper with odds at every fork.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;p&gt;
  &lt;a rel="noopener noreferrer" href="https://github.com/SrilalS/ForkPath/docs/logo.svg"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2FSrilalS%2FForkPath%2FHEAD%2Fdocs%2Flogo.svg" width="88" alt="Forkpath logo"&gt;&lt;/a&gt;
&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Forkpath&lt;/h1&gt;
&lt;/div&gt;

&lt;p&gt;
  &lt;b&gt;Change one moment in history. Watch the consequences branch out.&lt;/b&gt;&lt;br&gt;
  An AI writer maps what happens next, event by event, across parallel storylines,&lt;br&gt;
  with the odds at every fork. Runs on your machine; your timelines stay there
&lt;/p&gt;

&lt;p&gt;
  &lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/f6a4ee16ec8dab4413636510cb89396040ea9d683d24961987ad0187d929434c/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f7374617475732d616c7068612d663562303431"&gt;&lt;img alt="Status: alpha" src="https://camo.githubusercontent.com/f6a4ee16ec8dab4413636510cb89396040ea9d683d24961987ad0187d929434c/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f7374617475732d616c7068612d663562303431"&gt;&lt;/a&gt;
  &lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/48a3d8c8e3e59a6e4aa419f769e813ab7a2bf04d55af4661205cfa212e487eba/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f76657273696f6e2d302e312d386237636636"&gt;&lt;img alt="Version 0.1" src="https://camo.githubusercontent.com/48a3d8c8e3e59a6e4aa419f769e813ab7a2bf04d55af4661205cfa212e487eba/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f76657273696f6e2d302e312d386237636636"&gt;&lt;/a&gt;
  &lt;a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/70339a636aac084b74aee7abbe642e9514d6a38d341dd01bf30558c07c8853b2/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f656e67696e652d436c61756465253230436f6465253230253743253230616e792532306d6f64656c2532304150492d633963346666"&gt;&lt;img alt="Engines: Claude Code or any model API" src="https://camo.githubusercontent.com/70339a636aac084b74aee7abbe642e9514d6a38d341dd01bf30558c07c8853b2/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f656e67696e652d436c61756465253230436f6465253230253743253230616e792532306d6f64656c2532304150492d633963346666"&gt;&lt;/a&gt;
  &lt;a href="https://github.com/SrilalS/ForkPath/LICENSE" rel="noopener noreferrer"&gt;&lt;img alt="MIT license" src="https://camo.githubusercontent.com/4ee01516c3a2fbeebe63b3a93566f7655f5fc7d3a424d54937d05a13ed41e33a/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d346664316130"&gt;&lt;/a&gt;
&lt;/p&gt;

&lt;p&gt;
  &lt;a href="https://github.com/SrilalS/ForkPath/docs/guide.md" rel="noopener noreferrer"&gt;&lt;b&gt;Guide&lt;/b&gt;&lt;/a&gt; ·
  &lt;a href="https://github.com/SrilalS/ForkPath#getting-started" rel="noopener noreferrer"&gt;&lt;b&gt;Getting started&lt;/b&gt;&lt;/a&gt; ·
  &lt;a href="https://github.com/SrilalS/ForkPath/docs/engines.md" rel="noopener noreferrer"&gt;Engines and costs&lt;/a&gt; ·
  &lt;a href="https://github.com/SrilalS/ForkPath/docs/development.md" rel="noopener noreferrer"&gt;Development&lt;/a&gt;
&lt;/p&gt;

&lt;p&gt;&lt;a rel="noopener noreferrer" href="https://github.com/SrilalS/ForkPath/docs/screenshots/timeline.png"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2FSrilalS%2FForkPath%2FHEAD%2Fdocs%2Fscreenshots%2Ftimeline.png" alt="Forkpath mapping &amp;quot;The treasure fleets keep sailing&amp;quot;: two streams, forks with odds, and an event open in the side panel"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Why Forkpath&lt;/h2&gt;
&lt;/div&gt;

&lt;p&gt;Alternate history is usually told as one story. Real history is a tree of things that almost happened. Forkpath
keeps the whole tree:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;🌳 &lt;strong&gt;A map, not a story.&lt;/strong&gt; Every event is a node. The writer branches where outcomes are genuinely uncertain
gives each branch a likelihood, and tracks regions, wars and programs as parallel streams that influence and
merge into each other.&lt;/li&gt;
&lt;li&gt;⚖️ &lt;strong&gt;Neutral as water.&lt;/strong&gt; Language models lean toward tidy, peaceful endings. Forkpath's writer runs a realism
check (incentives, capabilities, historical base rates) before every event, and a separate review can flag
unearned outcomes, anachronisms and numbers out…&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/SrilalS/ForkPath" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;
&lt;br&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install
&lt;/span&gt;npm run dev
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then open &lt;a href="http://localhost:5173" rel="noopener noreferrer"&gt;http://localhost:5173&lt;/a&gt; and pick an engine in &lt;strong&gt;Settings&lt;/strong&gt;.&lt;/p&gt;

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

&lt;p&gt;TypeScript end to end: React 19 and React Flow (with dagre layout) in the browser, Express and Zod on the server, and the graph rules shared between the two. Timelines are plain JSON files on your disk.&lt;/p&gt;

&lt;p&gt;The model layer was designed to be swappable from day one:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;No SDKs.&lt;/strong&gt; The server talks to models through two small adapters built on plain &lt;code&gt;fetch&lt;/code&gt;: one for the Anthropic Messages API and one for any &lt;strong&gt;OpenAI-compatible&lt;/strong&gt; endpoint.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;That second adapter is the open door.&lt;/strong&gt; Point it at &lt;strong&gt;Ollama&lt;/strong&gt; (&lt;code&gt;http://localhost:11434/v1&lt;/code&gt;) or &lt;strong&gt;LM Studio&lt;/strong&gt; (&lt;code&gt;http://localhost:1234/v1&lt;/code&gt;) and an open-weight model running on your own machine writes your history. The Settings page has presets for both, plus a &lt;strong&gt;Test connection&lt;/strong&gt; button.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Same rules for every model.&lt;/strong&gt; Whatever engine you pick gets the same compact briefs, schema validation and writer rules, so swapping models doesn't mean rewriting prompts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Built to scale on a laptop.&lt;/strong&gt; Layout caching, viewport culling and a background layout worker keep thousand-node timelines smooth.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;Alternate history is personal. People use it to work through "what if" questions about their own country, family history or politics, and those are exactly the questions you might not want sitting in someone else's logs.&lt;/p&gt;

&lt;p&gt;With an open-weight model behind it, Forkpath is fully local:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🏠 &lt;strong&gt;Nothing leaves your machine.&lt;/strong&gt; The app, the model and your timelines all live on your laptop. There's no telemetry and no account.&lt;/li&gt;
&lt;li&gt;💸 &lt;strong&gt;Set it up once, never pay again.&lt;/strong&gt; A big timeline can mean hundreds of model calls. Over a paid API that adds up (Forkpath even has a monthly budget feature for that reason). A local model makes the meter disappear, so you can regenerate and branch freely.&lt;/li&gt;
&lt;li&gt;🔁 &lt;strong&gt;Swap the brain whenever you want.&lt;/strong&gt; A new open model comes out, you change one setting. Want a grimmer or more hopeful writer? Change the temperament, or try a different model and compare.&lt;/li&gt;
&lt;li&gt;📂 &lt;strong&gt;Open all the way down.&lt;/strong&gt; Forkpath itself is MIT-licensed, and timelines are plain JSON you can read, diff and share.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;Researching and writing a believable alternate timeline by hand takes days. In Forkpath it takes one sentence and a cup of tea. The writer handles the reasoning, the bookkeeping across storylines, and the "wait, would that actually happen?" check. You just decide which branch looks fun and click &lt;strong&gt;Continue&lt;/strong&gt;.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>opensource</category>
      <category>ai</category>
    </item>
    <item>
      <title>Scope Creep Guard - a local AI that tells freelancers when a client email is actually new work</title>
      <dc:creator>Srilal Sachintha</dc:creator>
      <pubDate>Fri, 16 Oct 2026 09:40:34 +0000</pubDate>
      <link>https://dev.to/srilals/scope-creep-guard-a-local-ai-that-tells-freelancers-when-a-client-email-is-actually-new-work-1i04</link>
      <guid>https://dev.to/srilals/scope-creep-guard-a-local-ai-that-tells-freelancers-when-a-client-email-is-actually-new-work-1i04</guid>
      <description>&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;Every freelancer knows this email:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Loving the designs so far! A few small things... could we also add online ordering? And a Spanish version? Oh, and could you write the About page copy?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Each line sounds small. Together they are a second project, and saying "that's extra" without sounding difficult is the hard part. Most of us either do the work for free or spend an hour re-reading the contract and writing a careful reply.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Scope Creep Guard&lt;/strong&gt; does that hour for you. You paste two things:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;📄 The scope you agreed (contract, proposal or SOW)&lt;/li&gt;
&lt;li&gt;✉️ The client's new email&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;It gives you back:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;✅ / ❌ / ❓ A verdict for every request: &lt;strong&gt;in scope&lt;/strong&gt;, &lt;strong&gt;out of scope&lt;/strong&gt; or &lt;strong&gt;unclear&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;🔖 The exact line of your contract that backs each verdict&lt;/li&gt;
&lt;li&gt;⏱️ An hour estimate for the extra work, priced at your hourly rate&lt;/li&gt;
&lt;li&gt;📨 A friendly, ready-to-send reply that confirms the in-scope work and lists the rest as a priced change order&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Who would pay for it:&lt;/strong&gt; freelance designers, developers, copywriters and small agencies. Anyone who sends proposals and gets "quick favour" emails. Unbilled scope creep easily costs a freelancer a few hundred dollars a month, so a one-time price of around $19 for a packaged desktop version pays for itself the first time it catches one extra page. The core stays open source and free to run.&lt;/p&gt;

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

&lt;p&gt;Here is a real run on the sample project in the repo: a 5-page bakery website, fixed price, two rounds of revisions, with copywriting, e-commerce and translations explicitly excluded. The client email asks for five things.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Request&lt;/th&gt;
&lt;th&gt;Verdict&lt;/th&gt;
&lt;th&gt;Contract line it cites&lt;/th&gt;
&lt;th&gt;Estimate&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Bigger menu text on mobile&lt;/td&gt;
&lt;td&gt;✅ In scope (a revision)&lt;/td&gt;
&lt;td&gt;-&lt;/td&gt;
&lt;td&gt;-&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pre-order birthday cakes with a deposit&lt;/td&gt;
&lt;td&gt;❌ Out of scope&lt;/td&gt;
&lt;td&gt;"Online ordering, payments or e-commerce of any kind."&lt;/td&gt;
&lt;td&gt;~4 h&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Spanish version of the site&lt;/td&gt;
&lt;td&gt;❌ Out of scope&lt;/td&gt;
&lt;td&gt;"Translations into other languages."&lt;/td&gt;
&lt;td&gt;~16 h&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A page for the second shop&lt;/td&gt;
&lt;td&gt;❌ Out of scope (6th page)&lt;/td&gt;
&lt;td&gt;-&lt;/td&gt;
&lt;td&gt;~4 h&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Write the "Our Story" paragraph&lt;/td&gt;
&lt;td&gt;❌ Out of scope&lt;/td&gt;
&lt;td&gt;"Copywriting and photography (client supplies all text and images)."&lt;/td&gt;
&lt;td&gt;~8 h&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;At $60/h that is &lt;strong&gt;USD 1,920 of work&lt;/strong&gt; that would otherwise have been done for free. The drafted reply:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Hi Maria,

Great to hear you're loving the designs! We'll definitely implement the changes
for the menu text on mobile. However, adding the pre-ordering functionality, the
Spanish translation, the additional shop page, and the 'Our Story' paragraph are
outside the original scope of this project. These can be added as a change order.

- pre-order birthday cakes on the site: ~4 h = USD 240.00
- Spanish version of the whole site: ~16 h = USD 960.00
- add a page for that shop too: ~4 h = USD 240.00
- write a short 'Our Story' paragraph: ~8 h = USD 480.00
Estimated total: USD 1,920.00

Best,
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&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/SrilalS" rel="noopener noreferrer"&gt;
        SrilalS
      &lt;/a&gt; / &lt;a href="https://github.com/SrilalS/scope-creep-guard" rel="noopener noreferrer"&gt;
        scope-creep-guard
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Local-first scope creep checker for freelancers: Gemma via Ollama flags out-of-scope client requests and drafts a priced change order.
    &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;🛡️ Scope Creep Guard&lt;/h1&gt;
&lt;/div&gt;

&lt;p&gt;Paste the scope you agreed with a client and their latest email. Scope Creep Guard flags which requests are
in scope, out of scope or unclear, quotes the line of your contract that backs each call, estimates the extra
hours, and drafts a polite reply with a priced change order.&lt;/p&gt;
&lt;p&gt;Everything runs on your own machine with an open-weight model (Gemma via Ollama). Your contracts and client
emails never leave your laptop.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Quick start&lt;/h2&gt;
&lt;/div&gt;
&lt;ol&gt;
&lt;li&gt;
&lt;p&gt;Install &lt;a href="https://ollama.com/download" rel="nofollow noopener noreferrer"&gt;Ollama&lt;/a&gt; and pull Gemma:&lt;/p&gt;
&lt;div class="highlight highlight-source-shell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;ollama pull gemma3:4b&lt;/pre&gt;

&lt;/div&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;Start the app (Python 3.10+, no extra packages needed):&lt;/p&gt;
&lt;div class="highlight highlight-source-shell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;python server.py&lt;/pre&gt;

&lt;/div&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;Open &lt;a href="http://localhost:8000" rel="nofollow noopener noreferrer"&gt;http://localhost:8000&lt;/a&gt;, click &lt;strong&gt;Load sample&lt;/strong&gt;, then &lt;strong&gt;Check for scope creep&lt;/strong&gt;.&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Configuration&lt;/h2&gt;
&lt;/div&gt;
&lt;div class="table-wrapper-paragraph"&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;SCG_MODEL&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;&lt;code&gt;gemma3:4b&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Any Ollama model, for example &lt;code&gt;gemma3:12b&lt;/code&gt; for better judgement&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;OLLAMA_URL&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;&lt;code&gt;http://127.0.0.1:11434&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Where Ollama is listening&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PORT&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;&lt;code&gt;8000&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Web server port&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;HOST&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;&lt;code&gt;127.0.0.1&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Bind address (local only by default)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;How it works&lt;/h2&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;scope_guard.py&lt;/code&gt; sends the…&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/SrilalS/scope-creep-guard" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;Three files do the work: &lt;code&gt;scope_guard.py&lt;/code&gt; (prompt, model call and pricing), &lt;code&gt;server.py&lt;/code&gt; (a tiny web server) and &lt;code&gt;static/index.html&lt;/code&gt; (the whole UI). No dependencies beyond Python and Ollama.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Stack:&lt;/strong&gt; &lt;a href="https://ollama.com/library/gemma3" rel="noopener noreferrer"&gt;Gemma 3 4B&lt;/a&gt; running locally through &lt;a href="https://ollama.com" rel="noopener noreferrer"&gt;Ollama&lt;/a&gt;, a Python standard-library web server, and one HTML page.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt; browser ──POST /api/analyze──▶ server.py ──/api/chat + JSON schema──▶ Ollama (gemma3:4b)
                                    │                                        │
                                    │◀────── verdicts + hour estimates ──────┘
                                    ▼
                         Python: price, build change order,
                         write summary, fill reply template
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;A few decisions made the difference with a small model:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Structured output, not "please reply in JSON".&lt;/strong&gt; Ollama accepts a JSON schema in the &lt;code&gt;format&lt;/code&gt; field and constrains generation to it. Gemma &lt;em&gt;has&lt;/em&gt; to return a list of items with &lt;code&gt;request&lt;/code&gt;, &lt;code&gt;verdict&lt;/code&gt; (an enum of the three values), &lt;code&gt;reason&lt;/code&gt;, &lt;code&gt;scope_quote&lt;/code&gt; and &lt;code&gt;est_hours&lt;/code&gt;. I never had to parse broken JSON.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. The model never does math.&lt;/strong&gt; My first version let the model write the prices into the email. Small models are bad at arithmetic, and a change order with a wrong total is worse than none. Now Gemma only classifies and estimates hours. Python multiplies by the rate, builds the list and drops it into the email where the model left a &lt;code&gt;{{CHANGE_ORDER}}&lt;/code&gt; token.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Don't trust the model to count either.&lt;/strong&gt; In one early run the summary said "three new requests" right above a list of five. The summary line is now computed from the verdicts ("4 of 5 requests go beyond the agreed scope").&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Teach it what a revision is.&lt;/strong&gt; On the first run Gemma marked "make the menu text bigger" as out of scope. Technically it was not listed, but any freelancer would call it a revision, and the contract includes two rounds. A short rule in the prompt fixed it: small changes to work that is already in scope are revisions, and included revisions make them in scope. It also learned that "add another page" to a 5-page contract is extra quantity.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Clean up after the model.&lt;/strong&gt; Gemma sometimes put the change-order token mid-sentence with a full stop after it. Rather than fight it with more prompt, a few lines of Python always give the table its own paragraph. These rules are covered by unit tests that run without a model.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Speed, honestly:&lt;/strong&gt; my laptop has no usable GPU, so &lt;code&gt;ollama ps&lt;/code&gt; shows the model at &lt;strong&gt;100% CPU&lt;/strong&gt; (2.9 GB loaded). One full check of the sample takes &lt;strong&gt;about 1 minute 45 seconds&lt;/strong&gt;. That's slow for a chat app but fine for this job: you paste the email, go make a coffee, and come back to a reply you would have spent an hour on. On a machine with a supported GPU, Ollama offloads the model and it should be much faster (I haven't measured that yet).&lt;/p&gt;
&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Because the input is a contract.&lt;/strong&gt; The two things this tool reads are the most confidential documents a freelancer has: a signed agreement (often under NDA) and a client's private email. Confidentiality clauses often limit sharing client material with third parties, and pasting it into someone else's cloud is exactly that. With an open-weight model on my own machine, nothing leaves the laptop. It even works offline, on a train, with the Wi-Fi off. That isn't a nice-to-have for this product; it's the reason a cautious freelancer would trust it with their contracts.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Because it makes the business model work.&lt;/strong&gt; A hosted API charges per call, forever, so a hosted version would need a subscription. Running Gemma locally makes every check free after the download. That's what makes a cheap, pay-once tool for solo freelancers possible.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Because I can swap the brain.&lt;/strong&gt; &lt;code&gt;SCG_MODEL=gemma3:12b python server.py&lt;/code&gt; and it uses a bigger Gemma for better judgement, or any other Ollama model. No vendor can deprecate the model under me or change its behaviour overnight, and a future version could be fine-tuned on real scope disputes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where closed would have been better:&lt;/strong&gt; speed. A hosted frontier model would answer in seconds and probably give tighter hour estimates. For a tool whose whole job is guarding private documents, I'll trade two minutes for keeping them on my laptop.&lt;/p&gt;
&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;I built this with Claude Code as my pair programmer: planning, writing the server and UI, running Gemma end to end and fixing the quality issues described above.## What I Built&lt;/p&gt;

&lt;p&gt;Every freelancer knows this email:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Loving the designs so far! A few small things... could we also add online ordering? And a Spanish version? Oh, and could you write the About page copy?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Each line sounds small. Together they are a second project, and saying "that's extra" without sounding difficult is the hard part. Most of us either do the work for free or spend an hour re-reading the contract and writing a careful reply.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Scope Creep Guard&lt;/strong&gt; does that hour for you. You paste two things:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;📄 The scope you agreed (contract, proposal or SOW)&lt;/li&gt;
&lt;li&gt;✉️ The client's new email&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;It gives you back:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;✅ / ❌ / ❓ A verdict for every request: &lt;strong&gt;in scope&lt;/strong&gt;, &lt;strong&gt;out of scope&lt;/strong&gt; or &lt;strong&gt;unclear&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;🔖 The exact line of your contract that backs each verdict&lt;/li&gt;
&lt;li&gt;⏱️ An hour estimate for the extra work, priced at your hourly rate&lt;/li&gt;
&lt;li&gt;📨 A friendly, ready-to-send reply that confirms the in-scope work and lists the rest as a priced change order&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Who would pay for it:&lt;/strong&gt; freelance designers, developers, copywriters and small agencies. Anyone who sends proposals and gets "quick favour" emails. Unbilled scope creep easily costs a freelancer a few hundred dollars a month, so a one-time price of around $19 for a packaged desktop version pays for itself the first time it catches one extra page. The core stays open source and free to run.&lt;/p&gt;
&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;



&lt;p&gt;Here is a real run on the sample project in the repo: a 5-page bakery website, fixed price, two rounds of revisions, with copywriting, e-commerce and translations explicitly excluded. The client email asks for five things.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Request&lt;/th&gt;
&lt;th&gt;Verdict&lt;/th&gt;
&lt;th&gt;Contract line it cites&lt;/th&gt;
&lt;th&gt;Estimate&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Bigger menu text on mobile&lt;/td&gt;
&lt;td&gt;✅ In scope (a revision)&lt;/td&gt;
&lt;td&gt;-&lt;/td&gt;
&lt;td&gt;-&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Pre-order birthday cakes with a deposit&lt;/td&gt;
&lt;td&gt;❌ Out of scope&lt;/td&gt;
&lt;td&gt;"Online ordering, payments or e-commerce of any kind."&lt;/td&gt;
&lt;td&gt;~4 h&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Spanish version of the site&lt;/td&gt;
&lt;td&gt;❌ Out of scope&lt;/td&gt;
&lt;td&gt;"Translations into other languages."&lt;/td&gt;
&lt;td&gt;~16 h&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A page for the second shop&lt;/td&gt;
&lt;td&gt;❌ Out of scope (6th page)&lt;/td&gt;
&lt;td&gt;-&lt;/td&gt;
&lt;td&gt;~4 h&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Write the "Our Story" paragraph&lt;/td&gt;
&lt;td&gt;❌ Out of scope&lt;/td&gt;
&lt;td&gt;"Copywriting and photography (client supplies all text and images)."&lt;/td&gt;
&lt;td&gt;~8 h&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;At $60/h that is &lt;strong&gt;USD 1,920 of work&lt;/strong&gt; that would otherwise have been done for free. The drafted reply:&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Hi Maria,

Great to hear you're loving the designs! We'll definitely implement the changes
for the menu text on mobile. However, adding the pre-ordering functionality, the
Spanish translation, the additional shop page, and the 'Our Story' paragraph are
outside the original scope of this project. These can be added as a change order.

- pre-order birthday cakes on the site: ~4 h = USD 240.00
- Spanish version of the whole site: ~16 h = USD 960.00
- add a page for that shop too: ~4 h = USD 240.00
- write a short 'Our Story' paragraph: ~8 h = USD 480.00
Estimated total: USD 1,920.00

Best,
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&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/SrilalS" rel="noopener noreferrer"&gt;
        SrilalS
      &lt;/a&gt; / &lt;a href="https://github.com/SrilalS/scope-creep-guard" rel="noopener noreferrer"&gt;
        scope-creep-guard
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Local-first scope creep checker for freelancers: Gemma via Ollama flags out-of-scope client requests and drafts a priced change order.
    &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;🛡️ Scope Creep Guard&lt;/h1&gt;
&lt;/div&gt;

&lt;p&gt;Paste the scope you agreed with a client and their latest email. Scope Creep Guard flags which requests are
in scope, out of scope or unclear, quotes the line of your contract that backs each call, estimates the extra
hours, and drafts a polite reply with a priced change order.&lt;/p&gt;
&lt;p&gt;Everything runs on your own machine with an open-weight model (Gemma via Ollama). Your contracts and client
emails never leave your laptop.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Quick start&lt;/h2&gt;
&lt;/div&gt;
&lt;ol&gt;
&lt;li&gt;
&lt;p&gt;Install &lt;a href="https://ollama.com/download" rel="nofollow noopener noreferrer"&gt;Ollama&lt;/a&gt; and pull Gemma:&lt;/p&gt;
&lt;div class="highlight highlight-source-shell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;ollama pull gemma3:4b&lt;/pre&gt;

&lt;/div&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;Start the app (Python 3.10+, no extra packages needed):&lt;/p&gt;
&lt;div class="highlight highlight-source-shell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;python server.py&lt;/pre&gt;

&lt;/div&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;Open &lt;a href="http://localhost:8000" rel="nofollow noopener noreferrer"&gt;http://localhost:8000&lt;/a&gt;, click &lt;strong&gt;Load sample&lt;/strong&gt;, then &lt;strong&gt;Check for scope creep&lt;/strong&gt;.&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Configuration&lt;/h2&gt;
&lt;/div&gt;
&lt;div class="table-wrapper-paragraph"&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;SCG_MODEL&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;&lt;code&gt;gemma3:4b&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Any Ollama model, for example &lt;code&gt;gemma3:12b&lt;/code&gt; for better judgement&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;OLLAMA_URL&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;&lt;code&gt;http://127.0.0.1:11434&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Where Ollama is listening&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PORT&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;&lt;code&gt;8000&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Web server port&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;HOST&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;&lt;code&gt;127.0.0.1&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Bind address (local only by default)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;How it works&lt;/h2&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;scope_guard.py&lt;/code&gt; sends the…&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/SrilalS/scope-creep-guard" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;Three files do the work: &lt;code&gt;scope_guard.py&lt;/code&gt; (prompt, model call and pricing), &lt;code&gt;server.py&lt;/code&gt; (a tiny web server) and &lt;code&gt;static/index.html&lt;/code&gt; (the whole UI). No dependencies beyond Python and Ollama.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Stack:&lt;/strong&gt; &lt;a href="https://ollama.com/library/gemma3" rel="noopener noreferrer"&gt;Gemma 3 4B&lt;/a&gt; running locally through &lt;a href="https://ollama.com" rel="noopener noreferrer"&gt;Ollama&lt;/a&gt;, a Python standard-library web server, and one HTML page.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt; browser ──POST /api/analyze──▶ server.py ──/api/chat + JSON schema──▶ Ollama (gemma3:4b)
                                    │                                        │
                                    │◀────── verdicts + hour estimates ──────┘
                                    ▼
                         Python: price, build change order,
                         write summary, fill reply template
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A few decisions made the difference with a small model:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Structured output, not "please reply in JSON".&lt;/strong&gt; Ollama accepts a JSON schema in the &lt;code&gt;format&lt;/code&gt; field and constrains generation to it. Gemma &lt;em&gt;has&lt;/em&gt; to return a list of items with &lt;code&gt;request&lt;/code&gt;, &lt;code&gt;verdict&lt;/code&gt; (an enum of the three values), &lt;code&gt;reason&lt;/code&gt;, &lt;code&gt;scope_quote&lt;/code&gt; and &lt;code&gt;est_hours&lt;/code&gt;. I never had to parse broken JSON.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. The model never does math.&lt;/strong&gt; My first version let the model write the prices into the email. Small models are bad at arithmetic, and a change order with a wrong total is worse than none. Now Gemma only classifies and estimates hours. Python multiplies by the rate, builds the list and drops it into the email where the model left a &lt;code&gt;{{CHANGE_ORDER}}&lt;/code&gt; token.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Don't trust the model to count either.&lt;/strong&gt; In one early run the summary said "three new requests" right above a list of five. The summary line is now computed from the verdicts ("4 of 5 requests go beyond the agreed scope").&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Teach it what a revision is.&lt;/strong&gt; On the first run Gemma marked "make the menu text bigger" as out of scope. Technically it was not listed, but any freelancer would call it a revision, and the contract includes two rounds. A short rule in the prompt fixed it: small changes to work that is already in scope are revisions, and included revisions make them in scope. It also learned that "add another page" to a 5-page contract is extra quantity.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Clean up after the model.&lt;/strong&gt; Gemma sometimes put the change-order token mid-sentence with a full stop after it. Rather than fight it with more prompt, a few lines of Python always give the table its own paragraph. These rules are covered by unit tests that run without a model.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Speed, honestly:&lt;/strong&gt; my laptop has no usable GPU, so &lt;code&gt;ollama ps&lt;/code&gt; shows the model at &lt;strong&gt;100% CPU&lt;/strong&gt; (2.9 GB loaded). One full check of the sample takes &lt;strong&gt;about 1 minute 45 seconds&lt;/strong&gt;. That's slow for a chat app but fine for this job: you paste the email, go make a coffee, and come back to a reply you would have spent an hour on. On a machine with a supported GPU, Ollama offloads the model and it should be much faster (I haven't measured that yet).&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Because the input is a contract.&lt;/strong&gt; The two things this tool reads are the most confidential documents a freelancer has: a signed agreement (often under NDA) and a client's private email. Confidentiality clauses often limit sharing client material with third parties, and pasting it into someone else's cloud is exactly that. With an open-weight model on my own machine, nothing leaves the laptop. It even works offline, on a train, with the Wi-Fi off. That isn't a nice-to-have for this product; it's the reason a cautious freelancer would trust it with their contracts.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Because it makes the business model work.&lt;/strong&gt; A hosted API charges per call, forever, so a hosted version would need a subscription. Running Gemma locally makes every check free after the download. That's what makes a cheap, pay-once tool for solo freelancers possible.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Because I can swap the brain.&lt;/strong&gt; &lt;code&gt;SCG_MODEL=gemma3:12b python server.py&lt;/code&gt; and it uses a bigger Gemma for better judgement, or any other Ollama model. No vendor can deprecate the model under me or change its behaviour overnight, and a future version could be fine-tuned on real scope disputes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where closed would have been better:&lt;/strong&gt; speed. A hosted frontier model would answer in seconds and probably give tighter hour estimates. For a tool whose whole job is guarding private documents, I'll trade two minutes for keeping them on my laptop.&lt;/p&gt;

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

&lt;p&gt;I built this with Claude Code as my pair programmer: planning, writing the server and UI, running Gemma end to end and fixing the quality issues described above.&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 4B, running fully locally through Ollama, is the core of the product.&lt;/li&gt;
&lt;/ul&gt;

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