<?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: DanialMo</title>
    <description>The latest articles on DEV Community by DanialMo (@danial_mo).</description>
    <link>https://dev.to/danial_mo</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%2F1896443%2F77d7caf4-1ddb-4fc2-ad62-e948cb3e4013.png</url>
      <title>DEV Community: DanialMo</title>
      <link>https://dev.to/danial_mo</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/danial_mo"/>
    <language>en</language>
    <item>
      <title>ChalkTalk: a whiteboard tutor that draws like a teacher, and lets you raise your hand</title>
      <dc:creator>DanialMo</dc:creator>
      <pubDate>Mon, 19 Oct 2026 06:55:52 +0000</pubDate>
      <link>https://dev.to/danial_mo/chalktalk-a-whiteboard-tutor-that-draws-like-a-teacher-and-lets-you-raise-your-hand-4jmd</link>
      <guid>https://dev.to/danial_mo/chalktalk-a-whiteboard-tutor-that-draws-like-a-teacher-and-lets-you-raise-your-hand-4jmd</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;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%2Fd9mbsahkm6zxqimg78ut.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%2Fd9mbsahkm6zxqimg78ut.png" alt=" " width="800" height="452"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Think about the best explanation you ever got. It probably wasn't a slide deck. It was someone with a marker, drawing the idea while they talked, and stopping when you said "wait, why?".&lt;/p&gt;

&lt;p&gt;When you prepare for an ML or system-design interview, you get the opposite: papers, long videos, and blog posts. None of them draw for you, and none of them let you raise your hand.&lt;/p&gt;

&lt;p&gt;So I built &lt;strong&gt;ChalkTalk&lt;/strong&gt;. You type a topic ("how attention works") or drop in a PDF, and a tutor teaches it on a whiteboard, out loud. It writes the example sentence word by word, draws the formula and circles the part it's talking about, builds a small table or bar chart, crosses out the wrong idea, and writes the key points at the end. When you're lost, you press ✋ (or &lt;code&gt;H&lt;/code&gt;). It finishes its sentence, answers on the side of the board while pointing at the drawing you're confused about, and then goes back to the lesson.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Three levels:&lt;/strong&gt; Basic, Intermediate, and Interview (with interview tips and follow-up questions).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Notes:&lt;/strong&gt; one short line per step appears next to the board, a summary, not a transcript.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;History:&lt;/strong&gt; replay any lesson or answer.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Voice on by default&lt;/strong&gt;, made inside your browser.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Who would pay?&lt;/strong&gt; People preparing for ML engineer, data scientist and system-design interviews already pay for courses, mock interviews and tutors. A tutor that teaches any topic or paper at 1 a.m., at your level, and takes questions, is the part they can't get today. The closest thing is Gauth's canvas mode, which is closed and general. ChalkTalk is narrow on purpose (ML and papers), and because the model is open and the voice runs on your device, a lesson costs me close to nothing. That means it can be cheap.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Try it:&lt;/strong&gt; &lt;a href="https://chalktalk-ohpg.onrender.com" rel="noopener noreferrer"&gt;https://chalktalk-ohpg.onrender.com&lt;/a&gt;, press &lt;strong&gt;Try as guest&lt;/strong&gt;, then &lt;strong&gt;Watch the demo&lt;/strong&gt;, or type any topic. It's on Render's free plan, so the first visit can take a minute to wake up.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F278l7jrlruvphnlt2m7b.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F278l7jrlruvphnlt2m7b.gif" alt=" " width="720" height="407"&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/Danialmoa" rel="noopener noreferrer"&gt;
        Danialmoa
      &lt;/a&gt; / &lt;a href="https://github.com/Danialmoa/chalktalk" rel="noopener noreferrer"&gt;
        chalktalk
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      A whiteboard tutor that talks and draws like a teacher. Raise your hand to ask questions.
    &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;ChalkTalk&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;A whiteboard tutor for ML and system-design interview prep. It talks and draws like a teacher at a board (the example sentence, formulas, small tables and bar charts, links, marks and crossed-out mistakes), with a pen dot that moves to what it's drawing. It writes the main points in a Notes panel as it goes, and you can raise your hand to ask a question at any time.&lt;/p&gt;
&lt;p&gt;Lessons are generated with &lt;strong&gt;Gemma 4&lt;/strong&gt; (&lt;code&gt;google/gemma-4-26b-a4b-it&lt;/code&gt;) through Backboard, which routes the call to OpenRouter. A small FastAPI server (&lt;code&gt;server/&lt;/code&gt;) keeps the Backboard key, handles log in and sign up, and serves the built page.&lt;/p&gt;
&lt;p&gt;Voice is Kokoro (&lt;code&gt;onnx-community/Kokoro-82M-v1.0-ONNX&lt;/code&gt;), running in the browser. It is on by default. The first visit downloads the model (about 115 MB); until it's ready, lessons play silently.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Demo&lt;/strong&gt; plays a built-in XGBoost lesson and needs no key. &lt;strong&gt;Try as&lt;/strong&gt;…&lt;/p&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/Danialmoa/chalktalk" 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;&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%2Finncum3w7aezo3gnk8dp.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%2Finncum3w7aezo3gnk8dp.png" alt=" " width="800" height="449"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Gemma 4 (26B-A4B, open weights)&lt;/strong&gt; writes every lesson and every answer, through Backboard, which routes the call to OpenRouter. &lt;strong&gt;Kokoro-82M&lt;/strong&gt;, an open voice model, speaks it inside your browser, in a Web Worker. &lt;strong&gt;Excalidraw&lt;/strong&gt;, the open-source whiteboard, is the board, so the hand-drawn look comes for free. A small &lt;strong&gt;FastAPI&lt;/strong&gt; app on &lt;strong&gt;Render&lt;/strong&gt; keeps the API key, handles accounts, and serves the page.&lt;/p&gt;

&lt;p&gt;One rule shaped everything: &lt;strong&gt;the model never gives pixels.&lt;/strong&gt; LLMs are bad at coordinates, and smaller open models are worse. So Gemma writes NDJSON, one step per line: what to say, a 3–8 word gist for the notes, and up to three drawing actions, placed with words like &lt;code&gt;below&lt;/code&gt; and &lt;code&gt;right_of&lt;/code&gt;. My code does the real layout.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"say"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"A mistake is the real price minus the guess."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"gist"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"Mistake = price − guess"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"actions"&lt;/span&gt;&lt;span class="p"&gt;:[{&lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"write"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"mistake"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"text"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"mistake = price − guess"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"at"&lt;/span&gt;&lt;span class="p"&gt;:{&lt;/span&gt;&lt;span class="nl"&gt;"region"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"main"&lt;/span&gt;&lt;span class="p"&gt;}},{&lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"mark"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"target"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"mistake.mistake"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="nl"&gt;"note"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="s2"&gt;"also called residual"&lt;/span&gt;&lt;span class="p"&gt;}]}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;One step per line also means streaming: step 1 is on the board while Gemma is still writing step 5.&lt;/p&gt;

&lt;p&gt;It worked on the first day. Then I watched it like a student, and it took four rounds to make it feel like a teacher.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. My first board looked like a flowchart, not a teacher
&lt;/h3&gt;

&lt;p&gt;The first version could write, draw boxes and arrows, and highlight. To keep the layout clean, I told Gemma "every item is a box". So every lesson became a chain of boxes: Query → Key → Value → Matching Score → Softmax… It even taught something wrong: Q, K and V are made side by side, not one from the other.&lt;/p&gt;

&lt;p&gt;A teacher draws the thing itself. So I gave the board a teacher's moves:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;words&lt;/code&gt;: a sentence split into words you can point at&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;link&lt;/code&gt;: a curved line between two words, thicker for a stronger link (attention weights, finally)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;table&lt;/code&gt;, &lt;code&gt;bars&lt;/code&gt;, and a &lt;code&gt;list&lt;/code&gt; that grows one line at a time&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;mark&lt;/code&gt;: underline or circle one part in red, with a short note under it&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;strike&lt;/code&gt;: cross out the wrong idea&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;And one small trick that made the biggest difference: any target can be &lt;code&gt;id.part&lt;/code&gt;. &lt;code&gt;s.cat&lt;/code&gt; is one word of a sentence, &lt;code&gt;houses.150k&lt;/code&gt; is one table cell, and &lt;code&gt;sum.η&lt;/code&gt; is the η inside a formula (I measure the text before it to find where to circle).&lt;/p&gt;

&lt;p&gt;The split is the important part: &lt;strong&gt;Gemma chooses &lt;em&gt;what&lt;/em&gt; to draw, the code decides &lt;em&gt;how it looks&lt;/em&gt; and &lt;em&gt;where it goes&lt;/em&gt;.&lt;/strong&gt; That keeps a 26B model reliable. The rest is staging: a new idea starts a new row, a tall board wraps into a second column like a teacher walking across the board, the camera moves to the spot &lt;em&gt;before&lt;/em&gt; drawing, a red pen dot follows whatever is being written, and at the end the camera zooms out to the whole board.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Getting reliable JSON out of Gemma
&lt;/h3&gt;

&lt;p&gt;The most confusing bug: "it stops after a couple of seconds". The lesson started, played two or three steps, and then sat on "Writing more steps…".&lt;/p&gt;

&lt;p&gt;The server logs were fine. The raw stream wasn't. Gemma closes almost every line with a stray &lt;code&gt;]}&lt;/code&gt;, like this: &lt;code&gt;…"note":"far away"}]}]}&lt;/code&gt;. &lt;code&gt;JSON.parse&lt;/code&gt; fails, and my parser quietly skipped the line. In one run, &lt;strong&gt;4 of the 5 opening steps disappeared&lt;/strong&gt;, so the lesson ran dry.&lt;/p&gt;

&lt;p&gt;The fix is small: read only the first complete JSON object on each line, counting brackets and ignoring anything inside strings. Two more lessons from the same hunt:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;One bad action no longer throws away its whole step and its spoken line.&lt;/li&gt;
&lt;li&gt;My own prompt listed "point at one part of an item…" as a bullet under Actions, and Gemma politely answered with &lt;code&gt;{"type":"point at"}&lt;/code&gt;. Small models take your formatting literally.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;After that: 10 of 10 steps valid in my test runs.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. The voice that froze the board, and then couldn't keep up
&lt;/h3&gt;

&lt;p&gt;Kokoro first ran on the page itself, and every sentence froze the board and the buttons. I measured the long tasks: 8.1 s, 6.1 s, 15.4 s. Moving it into a Web Worker fixed the freezing, but not the speed: it took &lt;strong&gt;10.7 s to make 7.3 s of speech&lt;/strong&gt;, so after a few long lines the voice fell behind and the lesson had gaps.&lt;/p&gt;

&lt;p&gt;The fix was two HTTP headers. onnxruntime-web, which runs Kokoro, only uses more than one thread when the page is &lt;em&gt;cross-origin isolated&lt;/em&gt; (it needs &lt;code&gt;SharedArrayBuffer&lt;/code&gt;):&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;response&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;headers&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Cross-Origin-Opener-Policy&lt;/span&gt;&lt;span class="sh"&gt;"&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;same-origin&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;span class="n"&gt;response&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;headers&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Cross-Origin-Embedder-Policy&lt;/span&gt;&lt;span class="sh"&gt;"&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;credentialless&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Same sentence, &lt;strong&gt;5.3 s&lt;/strong&gt;, now faster than real time. Safari ignores &lt;code&gt;credentialless&lt;/code&gt; and keeps one thread, and if the threaded version ever fails to load, the worker restarts on one thread, so the voice never just dies.&lt;/p&gt;

&lt;p&gt;The last gap was my own timing. Each step's drawing was spread over a duration &lt;em&gt;guessed&lt;/em&gt; from the text length, and the guess ran about 2 s long, so there was a silence after every line. Now the drawing follows the real length of the voice clip, and the voice is made three steps ahead, starting the moment a step arrives.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Before&lt;/th&gt;
&lt;th&gt;After&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Opening steps that survived parsing&lt;/td&gt;
&lt;td&gt;1 of 5&lt;/td&gt;
&lt;td&gt;5 of 5&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Time to make 7.3 s of speech&lt;/td&gt;
&lt;td&gt;10.7 s&lt;/td&gt;
&lt;td&gt;5.3 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Silence after each line&lt;/td&gt;
&lt;td&gt;~2 s&lt;/td&gt;
&lt;td&gt;~0 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Page freezes while speaking&lt;/td&gt;
&lt;td&gt;3–15 s&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;From Start to the first word&lt;/td&gt;
&lt;td&gt;(stalled)&lt;/td&gt;
&lt;td&gt;~7–10 s, with 5 steps ready&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  4. The bugs that would have ended my demo
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Invisible arrows.&lt;/strong&gt; Excalidraw's converter moves an arrow's first point if it isn't at (0, 0). Every arrow that pointed &lt;em&gt;up&lt;/em&gt; collapsed into a 1-pixel stub. Now every line is given relative to its first point.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The step that never ended.&lt;/strong&gt; On my Mac, Chrome's audio clock stopped while still reporting "running", so the clip's &lt;code&gt;ended&lt;/code&gt; event never fired and the lesson froze for good. Each clip now has a backup timer a little past its length.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pause moved the board.&lt;/strong&gt; Redrawing a step after pause or skip counted its own old spot as taken, so the item slid down a row every time.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Answers that stayed on the board.&lt;/strong&gt; Gemma sometimes placed an answer next to a lesson item, so "Continue" couldn't clear it. Anything drawn while answering is now forced into the side area.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Raise hand
&lt;/h3&gt;

&lt;p&gt;When you press ✋, the player finishes the current step and pauses. Gemma gets the lesson title, everything said so far, a one-line-per-item summary of the board (&lt;code&gt;id | type | text&lt;/code&gt;), and your question. It answers in the same NDJSON, in a side area, and it can mark items on the main board (&lt;code&gt;miss.C&lt;/code&gt;, circle, "worst"), so the answer points at the exact thing you didn't get.&lt;/p&gt;

&lt;h3&gt;
  
  
  Built with an AI agent, tested by measuring
&lt;/h3&gt;

&lt;p&gt;I built this with Claude Code as my pair. I played the student and said what felt wrong ("it looks like fixed components", "it seems stuck"), and the agent didn't guess: it drove the app in headless Chrome and measured long tasks, audio start and end times, and what Gemma actually sent. Every number in this post comes from those measurements. That's also how the two worst bugs were found: the stray brackets and the stuck audio clock.&lt;/p&gt;

&lt;p&gt;The backend is small on purpose: FastAPI with email and password accounts (scrypt hashes, a signed session cookie), Postgres on Render, and an &lt;code&gt;/api/llm&lt;/code&gt; proxy that picks the model, caps the size, and rate-limits each client, so the Backboard key never reaches the browser.&lt;/p&gt;

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

&lt;p&gt;For a tutor, the voice and the model are the product, and that's exactly where open matters.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;It's what makes it cheap.&lt;/strong&gt; The voice is made on your device, so there's no per-character bill. A lesson is a minute or two of speech, and with a paid voice API the margin is gone before you've asked a question.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It's private.&lt;/strong&gt; The voice never leaves your computer, and history stays in your browser.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I'm not locked in.&lt;/strong&gt; Gemma's weights are open. Today it runs through Backboard; tomorrow it can run on my own GPU, or be fine-tuned on how good teachers explain.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I could read the code.&lt;/strong&gt; I fixed the invisible arrows by reading Excalidraw's converter, and doubled the voice speed by learning how onnxruntime-web picks threads. With a closed API, both would have been a support ticket.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;ul&gt;
&lt;li&gt;Real stroke-by-stroke handwriting instead of fade-ins&lt;/li&gt;
&lt;li&gt;Formulas rendered properly (KaTeX) inside the board&lt;/li&gt;
&lt;li&gt;Ask out loud: hold Space to talk, with Whisper in the browser&lt;/li&gt;
&lt;li&gt;History saved per account, not per browser&lt;/li&gt;
&lt;/ul&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 4 writes every lesson and every answer, and decides what the board draws.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render:&lt;/strong&gt; the whole app, the FastAPI backend, the built page and Postgres, runs on Render from one &lt;code&gt;render.yaml&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Backboard:&lt;/strong&gt; one Backboard key streams Gemma 4 for lessons and raise-hand answers, kept on the server behind a rate-limited proxy.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Spoke &amp; Story : a bike route planner with a storyteller in your ear</title>
      <dc:creator>DanialMo</dc:creator>
      <pubDate>Sun, 11 Oct 2026 14:10:05 +0000</pubDate>
      <link>https://dev.to/danial_mo/spoke-story-55j7</link>
      <guid>https://dev.to/danial_mo/spoke-story-55j7</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week1-2026-10-05"&gt;Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;&lt;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%2Free7ggyu4bltu5mrysab.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%2Free7ggyu4bltu5mrysab.png" alt=" " width="72" height="72"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;One of the reasons I don't go out on my bike is that I don't know where to go. I want somewhere beautiful, without high hills. I'd also take the bike to the gym or anywhere else if I knew the best way there. And riding is more fun when you can listen to something: stories about the places you pass, a bit of news, like old-school radio :))&lt;/p&gt;

&lt;p&gt;So I built &lt;strong&gt;Spoke &amp;amp; Story&lt;/strong&gt;. You tell it where you want to go, or just "something flat, my legs are dead". It finds a nice bike route, and while you ride, Santa tells you one silly, connected story about the places you pass. The story is built from real Wikipedia facts, and each chapter is written live as you get close to the place.&lt;/p&gt;

&lt;p&gt;It's for anyone who has a bike in the garage and no good reason to take it out.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Where to?&lt;/strong&gt; Type a place, pick Scenic, Flattest, Fastest or Bike lanes. You get up to 3 routes, and Gemma picks one and tells you why.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ride buddy:&lt;/strong&gt; prefer to chat? An agent asks at most two questions, then plans the ride.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ride mode:&lt;/strong&gt; follows your GPS and gives you the real turns only. Santa tells a chapter when you get near a place, talks about the town when it's quiet, and finds a new route when you go off course.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Remember me:&lt;/strong&gt; ride the same route again and the story continues as the next episode. Everything personal stays on your phone.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;&lt;strong&gt;Try it:&lt;/strong&gt; &lt;a href="https://spoke-and-story.tech" rel="noopener noreferrer"&gt;https://spoke-and-story.tech&lt;/a&gt; (best on a phone. At home? Press &lt;strong&gt;Simulate&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/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fry7qxsq7c34lcbshii8n.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fry7qxsq7c34lcbshii8n.gif" alt=" " width="393" height="852"&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/Danialmoa" rel="noopener noreferrer"&gt;
        Danialmoa
      &lt;/a&gt; / &lt;a href="https://github.com/Danialmoa/spoke-and-story" rel="noopener noreferrer"&gt;
        spoke-and-story
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Bike route planner with a funny storytelling ride guide (Gemma 4, Kokoro, OpenStreetMap)
    &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;Spoke &amp;amp; Story&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;A bike route planner with a funny ride guide. Tell it where you want to go, it finds a nice route, and on the way
Santa tells you one silly, connected story about the places you pass.&lt;/p&gt;
&lt;p&gt;Built for the DEV Hacktoberfest 2026 "Touch Grass" challenge: open-source AI that gets you outside.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Try it:&lt;/strong&gt; &lt;a href="https://spoke-and-story.tech" rel="nofollow noopener noreferrer"&gt;https://spoke-and-story.tech&lt;/a&gt; (best on a phone: ride mode uses your GPS)&lt;/p&gt;
&lt;p&gt;
  &lt;a rel="noopener noreferrer" href="https://github.com/Danialmoa/spoke-and-story/.github/media/demo.gif"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fraw.githubusercontent.com%2FDanialmoa%2Fspoke-and-story%2FHEAD%2F.github%2Fmedia%2Fdemo.gif" width="300" alt="10-second demo on a phone: typing Como, three routes on the map, Gemma's pick, then a simulated ride where Santa starts the story"&gt;&lt;/a&gt;
&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What you can do&lt;/h2&gt;
&lt;/div&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Where to?&lt;/strong&gt; Type a destination (typos are fine), pick Scenic, Flattest or Fastest and your bike. You get up to
3 routes in a few seconds, then Gemma picks one and tells you why.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ride buddy:&lt;/strong&gt; prefer to chat? Say what you feel like ("50 km gravel loop", "something flat, my legs are dead")
It asks at most two questions, then plans the ride.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Saved places and Home:&lt;/strong&gt; tap ☆ next to the destination to save it, and set…&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/Danialmoa/spoke-and-story" 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;&lt;strong&gt;Gemma 4 (26B-A4B, open weights) is the brain.&lt;/strong&gt; It picks the route and explains why. It runs the ride buddy, a small agent that can search places and plan routes. It writes the story outline and then each chapter live, from the Wikipedia facts about the place you're riding past. &lt;strong&gt;Kokoro-82M&lt;/strong&gt;, an open-source voice model, can run right inside your phone's browser. The map side is all open too: OpenStreetMap data through Photon (place search), OpenRouteService (routing) and Overpass (water, forest, viewpoints, cafés).&lt;/p&gt;

&lt;p&gt;It took me three tries.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;First try:&lt;/strong&gt; a route planner. Type a place, get three routes, Gemma picks one, and Santa reads a story written before the ride, in Kokoro's voice. It worked well at my desk.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Second try:&lt;/strong&gt; I put it on Render and took it outside, and that's where it broke. First, my iPhone made no sound at all. Then on my first real ride, the app said "44 lines ready"... and Santa said nothing. Every story was tied to a kilometre on the planned route, and each time I took another road, the reroute threw them away. So I changed the idea: every place gets a circle around it, and when you ride into the circle, Gemma writes that chapter right then. When it's quiet for a while, Santa tells you something about the town you're in.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Third try:&lt;/strong&gt; Kokoro was too slow on my phone, so ElevenLabs became the default voice, with Kokoro as the open-source option (it now speaks sentence by sentence). Stories are made little by little, a few km ahead, instead of all at the start. Gemma also suggests topics about your town (like the history of the train in Bergamo), and each one is checked against Wikipedia before Santa says it, so he doesn't make up history. Ride the same route twice and you get the next episode, starting from the last ride's cliffhanger.&lt;/p&gt;

&lt;p&gt;And it actually works: it got me out a couple of times to test it for real. I encourage you to go outside to try it, but I added a simulation too, so you can see how it works from home :))&lt;/p&gt;

&lt;p&gt;I shared the build as it happened:&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2108602757246541868-87" src="https://platform.twitter.com/embed/Tweet.html?id=2108602757246541868"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2108602757246541868-87');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2108602757246541868&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2109218495871484211-842" src="https://platform.twitter.com/embed/Tweet.html?id=2109218495871484211"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2109218495871484211-842');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2109218495871484211&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;The backend is a small FastAPI app, the page is plain HTML and JavaScript with no build step, and it all runs on Render.&lt;/p&gt;
&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;



&lt;p&gt;It matters because open models can run without limits, and locally, without sharing a lot of your data. After the first try, I switched the voice to ElevenLabs, but I believe this part can run on open weights too. I just need more time to improve it.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Gemma's weights are open, so I'm not locked to one provider. It already runs through two hosts (Backboard, with Google AI Studio as the fallback), and I could move it to my own machine.&lt;/li&gt;
&lt;li&gt;Kokoro runs inside your phone's browser, so with that voice nothing you hear leaves your phone.&lt;/li&gt;
&lt;li&gt;Routes, places and facts all come from open data (OpenStreetMap, Wikipedia), which I can query and score freely, like counting the water and forest along a route.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&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 4 is the agent at the core: the route pick, the ride buddy, the live stories and the topics.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Render:&lt;/strong&gt; the whole app (the agent's backend and the front end) runs on Render at spoke-and-story.tech.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of ElevenLabs:&lt;/strong&gt; ElevenLabs gives the open-source Gemma agent its voice, with the costs capped on the server.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Backboard:&lt;/strong&gt; one Backboard API key serves Gemma 4, in about 2 s a call instead of 20-40 s.&lt;/li&gt;
&lt;/ul&gt;

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