<?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: Hyorinmaru</title>
    <description>The latest articles on DEV Community by Hyorinmaru (@hyorinmaruk).</description>
    <link>https://dev.to/hyorinmaruk</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%2F4216024%2F5095dffa-6945-4a17-bdd4-0cdaed4dddc1.png</url>
      <title>DEV Community: Hyorinmaru</title>
      <link>https://dev.to/hyorinmaruk</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/hyorinmaruk"/>
    <language>en</language>
    <item>
      <title>How I Parse Recipe Ingredients in Six Languages Without a Server</title>
      <dc:creator>Hyorinmaru</dc:creator>
      <pubDate>Tue, 10 Nov 2026 15:07:29 +0000</pubDate>
      <link>https://dev.to/hyorinmaruk/how-i-parse-recipe-ingredients-in-six-languages-without-a-server-4c38</link>
      <guid>https://dev.to/hyorinmaruk/how-i-parse-recipe-ingredients-in-six-languages-without-a-server-4c38</guid>
      <description>&lt;p&gt;Kerfuffle is the recipe app I built because I wanted my recipes on my phone, and nowhere else. No account, no server, no ads, no analytics. I built it solo, in Flutter. With no server to call, a shared recipe or a typed shopping-list line has to be parsed on the phone.&lt;/p&gt;

&lt;p&gt;The app speaks English, Spanish, Japanese, Kazakh, Korean, and Russian. A shopping list for the week has to be able to merge &lt;code&gt;2 cups flour&lt;/code&gt; with &lt;code&gt;2 стакана муки&lt;/code&gt; and &lt;code&gt;밀가루 2컵&lt;/code&gt; into one line for flour. A model on a server would have broken the offline rule. An on-device model felt like a lot of machinery for a line that is usually just a number, a unit, and a name, written in a handful of word orders.&lt;/p&gt;

&lt;p&gt;So the parser is rule-based. It is a pure function: one raw line in, a quantity, a display string, a unit, a name, and an optional note out. It runs offline on every import line. And when a line is ambiguous, it prefers no quantity over a wrong guess. Imports land as a draft you review before saving, so a blank quantity is easy to fix. A confidently wrong one is not.&lt;/p&gt;

&lt;h2&gt;
  
  
  Four grammars, first match wins
&lt;/h2&gt;

&lt;p&gt;There is no separate tokenizer, and the parser does not take a language code. One combined table covers all six languages. Every line goes through the same steps.&lt;/p&gt;

&lt;p&gt;First the line is normalized. Non-breaking spaces and tabs become ordinary spaces, a leading bullet is stripped, and extra whitespace is collapsed. Then notes are peeled off before any grammar runs: parentheses, including full-width &lt;code&gt;（）&lt;/code&gt;, and "to taste"-style phrases in all six languages — &lt;code&gt;to taste&lt;/code&gt;, &lt;code&gt;al gusto&lt;/code&gt;, &lt;code&gt;по вкусу&lt;/code&gt;, &lt;code&gt;дәміне қарай&lt;/code&gt;, &lt;code&gt;약간&lt;/code&gt; / &lt;code&gt;적당량&lt;/code&gt;, &lt;code&gt;適量&lt;/code&gt; / &lt;code&gt;少々&lt;/code&gt;, and a few synonyms.&lt;/p&gt;

&lt;p&gt;What remains is tried against four grammars, in order. The first match wins:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight dart"&gt;&lt;code&gt;&lt;span class="c1"&gt;/// Grammar, tried in this order:&lt;/span&gt;
&lt;span class="c1"&gt;///   1. `[qty] [unit] [of] name [, note]`      — en / ru / kk, and "2컵 밀가루"&lt;/span&gt;
&lt;span class="c1"&gt;///   2. `пол(-)unit name`                       — Russian "полстакана"&lt;/span&gt;
&lt;span class="c1"&gt;///   3. `unit [of] name` (quantity = 1)         — "pinch of salt", "щепотка соли"&lt;/span&gt;
&lt;span class="c1"&gt;///   4. `name [—-:] qty [unit] [note]`          — Korean default, Russian sites&lt;/span&gt;
&lt;span class="c1"&gt;///   5. name only.&lt;/span&gt;
&lt;span class="c1"&gt;/// Parenthesised text and "to taste"-style phrases are pulled into `note`&lt;/span&gt;
&lt;span class="c1"&gt;/// before any of that runs.&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight dart"&gt;&lt;code&gt;    &lt;span class="kd"&gt;final&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;_parseLeading&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt;
        &lt;span class="nf"&gt;_parsePolPrefix&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt;
        &lt;span class="nf"&gt;_parseUnitFirst&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt;
        &lt;span class="nf"&gt;_parseTrailing&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Leading quantity covers English and Spanish (&lt;code&gt;2 cups flour&lt;/code&gt;, &lt;code&gt;2 tazas de harina&lt;/code&gt;), Russian and Kazakh, and Korean when the amount comes first (&lt;code&gt;2컵 밀가루&lt;/code&gt;). If &lt;code&gt;dozen&lt;/code&gt; follows the number, the quantity becomes twelve and the unit is cleared, so &lt;code&gt;1 dozen eggs&lt;/code&gt; is 12 eggs, not one dozen.&lt;/p&gt;

&lt;p&gt;The Russian &lt;code&gt;пол-&lt;/code&gt; prefix is its own rule, because &lt;code&gt;полстакана молока&lt;/code&gt; and &lt;code&gt;пол чайной ложки соли&lt;/code&gt; mean half, and that does not belong in the generic number-word table. &lt;code&gt;Полба&lt;/code&gt;, the grain, is explicitly not "half of ба".&lt;/p&gt;

&lt;p&gt;Unit-first lines have no number. &lt;code&gt;pinch of salt&lt;/code&gt; and &lt;code&gt;щепотка соли&lt;/code&gt; both become quantity 1.&lt;/p&gt;

&lt;p&gt;The trailing rule is the default for Korean (&lt;code&gt;밀가루 2컵&lt;/code&gt;) and is common on Russian and Kazakh sites (&lt;code&gt;Мука — 200 г&lt;/code&gt;, &lt;code&gt;Ұн — 2 стакан&lt;/code&gt;). Japanese spoons write the unit &lt;em&gt;before&lt;/em&gt; the number (&lt;code&gt;醤油 大さじ3&lt;/code&gt;), so after the trailing parser finds a quantity it looks up to eight characters back for a unit that ends exactly there.&lt;/p&gt;

&lt;p&gt;If nothing matches, the cleaned line is the name, and quantity and unit stay empty. That is how &lt;code&gt;Salt to taste&lt;/code&gt;, &lt;code&gt;소금 약간&lt;/code&gt;, and &lt;code&gt;塩 適量&lt;/code&gt; are stored: a name and a note, no invented amount.&lt;/p&gt;

&lt;p&gt;Units themselves come from a pattern table. At a given position the longest match wins, so &lt;code&gt;ст. л.&lt;/code&gt; beats &lt;code&gt;ст.&lt;/code&gt; and &lt;code&gt;큰술&lt;/code&gt; beats &lt;code&gt;큰&lt;/code&gt;. Case-sensitive single-letter spoons (&lt;code&gt;1 T&lt;/code&gt; for a tablespoon, &lt;code&gt;1 t&lt;/code&gt; for a teaspoon) sit outside the table and only count immediately after a quantity, so &lt;code&gt;T-bone steak&lt;/code&gt; is not a tablespoon of steak.&lt;/p&gt;

&lt;h2&gt;
  
  
  A cup is not always 240 ml
&lt;/h2&gt;

&lt;p&gt;Regional cups are different units, not one unit with a locale flag, because conversion needs a different millilitre value for each:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight dart"&gt;&lt;code&gt;    &lt;span class="s1"&gt;'cup'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;UnitDef&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'cup'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitKind&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;volume&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitSystem&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;us&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;240&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s1"&gt;'cup_ru'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;UnitDef&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'cup_ru'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitKind&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;volume&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitSystem&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ru&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;250&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s1"&gt;'cup_ko'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;UnitDef&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'cup_ko'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitKind&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;volume&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitSystem&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ko&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;200&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s1"&gt;'cup_ja'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;UnitDef&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'cup_ja'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitKind&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;volume&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitSystem&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ja&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;200&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s1"&gt;'go'&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;UnitDef&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'go'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitKind&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;volume&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;UnitSystem&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ja&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;180&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A US cup is 240 ml. A Russian or Kazakh &lt;code&gt;стакан&lt;/code&gt; / &lt;code&gt;стақан&lt;/code&gt; is 250 ml. A Korean &lt;code&gt;컵&lt;/code&gt; and a Japanese &lt;code&gt;カップ&lt;/code&gt; are 200 ml. Japanese &lt;code&gt;合&lt;/code&gt;, used for rice, is 180 ml. Korean &lt;code&gt;근&lt;/code&gt; is a mass unit, 600 g. Tablespoons and teaspoons are 15 ml and 5 ml everywhere.&lt;/p&gt;

&lt;p&gt;Volume and mass convert through those bases. Count units never do: pieces, cloves, pinches, cans, bunches, and the rest stay as counts. If a conversion is refused, the original value and unit are left alone.&lt;/p&gt;

&lt;h2&gt;
  
  
  Fractions, decimal commas, and ranges
&lt;/h2&gt;

&lt;p&gt;Numeric amounts are matched longest-first, so &lt;code&gt;1 1/2&lt;/code&gt; is not read as &lt;code&gt;1&lt;/code&gt;. Mixed numbers work with &lt;code&gt;and&lt;/code&gt;, Russian &lt;code&gt;и&lt;/code&gt;, Kazakh &lt;code&gt;және&lt;/code&gt;, and Korean &lt;code&gt;과&lt;/code&gt; (&lt;code&gt;양파 1과 1/2개&lt;/code&gt;). Unicode fractions live in a lookup table — &lt;code&gt;½&lt;/code&gt;, &lt;code&gt;¼&lt;/code&gt;, &lt;code&gt;1½&lt;/code&gt;, and the rest of the usual set — and so do ASCII fractions like &lt;code&gt;1/2&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Decimals accept a dot or a comma: &lt;code&gt;1.5&lt;/code&gt;, &lt;code&gt;0,5&lt;/code&gt;, &lt;code&gt;1,5 кг&lt;/code&gt;. The comma form is what Russian and Kazakh recipes use.&lt;/p&gt;

&lt;p&gt;Ranges keep the original text in the display field and use the lower bound as the numeric quantity. Separators are &lt;code&gt;-&lt;/code&gt;, &lt;code&gt;–&lt;/code&gt;, &lt;code&gt;—&lt;/code&gt;, &lt;code&gt;~&lt;/code&gt;, and the words &lt;code&gt;to&lt;/code&gt;, &lt;code&gt;до&lt;/code&gt;, &lt;code&gt;или&lt;/code&gt;, &lt;code&gt;немесе&lt;/code&gt;, and &lt;code&gt;또는&lt;/code&gt;. &lt;code&gt;2-3&lt;/code&gt;, &lt;code&gt;2 to 3&lt;/code&gt;, &lt;code&gt;10–12&lt;/code&gt;, and &lt;code&gt;청양고추 1~2개&lt;/code&gt; all parse that way.&lt;/p&gt;

&lt;p&gt;A percent sign after a number is rejected on purpose. &lt;code&gt;2% milk&lt;/code&gt; and &lt;code&gt;Молоко 3,2%&lt;/code&gt; stay names. &lt;code&gt;1 cup 2% milk&lt;/code&gt; keeps &lt;code&gt;2% milk&lt;/code&gt; as the name. &lt;code&gt;Сметана 20% — 200 г&lt;/code&gt; keeps the fat percentage in the name and takes 200 g as the amount.&lt;/p&gt;

&lt;h2&gt;
  
  
  The lines that actually show up
&lt;/h2&gt;

&lt;p&gt;Most of the work was ordinary recipe phrasing, not theoretically messy input.&lt;/p&gt;

&lt;p&gt;In English, &lt;code&gt;1 (14 oz) can diced tomatoes&lt;/code&gt; is quantity 1, unit &lt;code&gt;can&lt;/code&gt;, name &lt;code&gt;diced tomatoes&lt;/code&gt;, note &lt;code&gt;14 oz&lt;/code&gt;. Multiple notes join: &lt;code&gt;1 medium onion (about 150 g), diced&lt;/code&gt;. &lt;code&gt;Juice of 1 lemon&lt;/code&gt; and &lt;code&gt;Zest of one lemon&lt;/code&gt; stay whole names; the trailing rule will not split them into a name and a bare number. Bullets are stripped, so &lt;code&gt;- 2 cups flour&lt;/code&gt; and &lt;code&gt;• 2 cups flour&lt;/code&gt; are the same line.&lt;/p&gt;

&lt;p&gt;Russian sites like to put the name first: &lt;code&gt;Мука — 2 стакана&lt;/code&gt;, &lt;code&gt;Яйца 3 шт&lt;/code&gt;, &lt;code&gt;Лук 1 шт крупный&lt;/code&gt; (the note is &lt;code&gt;крупный&lt;/code&gt;). Abbreviations such as &lt;code&gt;0,5 ч.л. соды&lt;/code&gt; and &lt;code&gt;ст. л.&lt;/code&gt; have to be the longest match, as above. Unit-first also shows up: &lt;code&gt;щепотка соли&lt;/code&gt;, &lt;code&gt;пучок укропа&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Kazakh recipes use their own units — &lt;code&gt;ас қасық&lt;/code&gt;, &lt;code&gt;шай қасық&lt;/code&gt;, &lt;code&gt;келі&lt;/code&gt;, &lt;code&gt;тіс&lt;/code&gt;, &lt;code&gt;бау&lt;/code&gt;, &lt;code&gt;уыс&lt;/code&gt;, &lt;code&gt;шымшым&lt;/code&gt; — and they also use Russian units freely. &lt;code&gt;2 стакан ұн&lt;/code&gt; is still a 250 ml &lt;code&gt;стакан&lt;/code&gt;. Number words work (&lt;code&gt;жарты стакан сүт&lt;/code&gt;, &lt;code&gt;екі жұмыртқа&lt;/code&gt;), and so does name-first order (&lt;code&gt;Қант 100 г&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;Korean defaults to name, then amount and counter, often with no space: &lt;code&gt;밀가루 2컵&lt;/code&gt;, &lt;code&gt;달걀 2개&lt;/code&gt;, &lt;code&gt;두부 1모&lt;/code&gt;, &lt;code&gt;간장 2큰술&lt;/code&gt;, &lt;code&gt;고기 1근&lt;/code&gt;. Native number words attach to the counter (&lt;code&gt;계란 한 개&lt;/code&gt;, &lt;code&gt;양파 반개&lt;/code&gt;). &lt;code&gt;두세 개 마늘&lt;/code&gt; is an approximate range and becomes 2. The lookbehind on the quantity regex only blocks Latin and Cyrillic letters, not Hangul, kana, or kanji, which is why &lt;code&gt;2컵&lt;/code&gt; and &lt;code&gt;大さじ3&lt;/code&gt; match at all:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight dart"&gt;&lt;code&gt;    &lt;span class="c1"&gt;// A digit must not continue a word ("abc3", "тест3"), but Japanese and&lt;/span&gt;
    &lt;span class="c1"&gt;// Korean glue the counter straight onto the number ("大さじ3", "2컵"),&lt;/span&gt;
    &lt;span class="c1"&gt;// so only Latin and Cyrillic letters block a match.&lt;/span&gt;
    &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="s1"&gt;'(?&amp;lt;![\d.,/])(?&amp;lt;![A-Za-zЀ-ӿҠ-һ])'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;한&lt;/code&gt; is also the word for "one", but &lt;code&gt;한우 300g&lt;/code&gt; has to stay Korean beef plus 300 g, not "one 우" plus 300 g.&lt;/p&gt;

&lt;p&gt;Japanese is name then number and counter (&lt;code&gt;小麦粉 2カップ&lt;/code&gt;, &lt;code&gt;米 2合&lt;/code&gt;, &lt;code&gt;豆腐 1/2丁&lt;/code&gt;), except for those spoons. Full-width parentheses become notes: &lt;code&gt;玉ねぎ 1個(薄切り)&lt;/code&gt;. The "to taste" pattern for &lt;code&gt;少々&lt;/code&gt; is bounded so it does not swallow &lt;code&gt;こしょう&lt;/code&gt; out of &lt;code&gt;黒こしょう&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Spanish strips &lt;code&gt;de&lt;/code&gt; / &lt;code&gt;del&lt;/code&gt; the way English strips &lt;code&gt;of&lt;/code&gt;, but it cannot stop at the first &lt;code&gt;de&lt;/code&gt;. &lt;code&gt;3 cdas de salsa de soja&lt;/code&gt; has to remain soy sauce, not "salsa".&lt;/p&gt;

&lt;h2&gt;
  
  
  One flour, six names
&lt;/h2&gt;

&lt;p&gt;Parsing is only half of the shopping list. After a line is parsed, &lt;code&gt;IngredientMatcher&lt;/code&gt; maps the name to a canonical id and an aisle. The dictionary ships with the app: 144 ingredients, aliases in all six languages, ten aisles. &lt;code&gt;flour&lt;/code&gt;, &lt;code&gt;мука&lt;/code&gt;, &lt;code&gt;밀가루&lt;/code&gt;, &lt;code&gt;小麦粉&lt;/code&gt;, &lt;code&gt;ұн&lt;/code&gt;, and &lt;code&gt;harina&lt;/code&gt; share one id. Prep words are stripped (&lt;code&gt;finely chopped onion&lt;/code&gt;, &lt;code&gt;cebolla picada&lt;/code&gt;, &lt;code&gt;다진 마늘&lt;/code&gt;, &lt;code&gt;おろしにんにく&lt;/code&gt;), but a name whose identity includes a prep word is tried exact-first, so &lt;code&gt;whole chicken&lt;/code&gt; is not &lt;code&gt;chicken breast&lt;/code&gt;, while &lt;code&gt;whole milk&lt;/code&gt; still becomes milk. The longest alias wins: &lt;code&gt;sesame oil&lt;/code&gt; is not &lt;code&gt;oil&lt;/code&gt;. English and Spanish plurals are handled carefully enough that a naïve &lt;code&gt;es&lt;/code&gt; strip cannot turn &lt;code&gt;cloves&lt;/code&gt; into &lt;code&gt;clov&lt;/code&gt;. Russian genitives are dictionary aliases (&lt;code&gt;муки&lt;/code&gt; → &lt;code&gt;мука&lt;/code&gt;), not stemming. Green onion, scallions, &lt;code&gt;зелёного лука&lt;/code&gt;, &lt;code&gt;파&lt;/code&gt;, &lt;code&gt;cebolletas&lt;/code&gt;, and &lt;code&gt;青ねぎ&lt;/code&gt; share one id. Unknown names still get a stable lowercase key and land in the "other" aisle, so two copies of the same unknown still merge with each other.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;GroceryMerger&lt;/code&gt; then folds lines that share a canonical id, with a few hard rules:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight dart"&gt;&lt;code&gt;  &lt;span class="k"&gt;static&lt;/span&gt; &lt;span class="kt"&gt;bool&lt;/span&gt; &lt;span class="nf"&gt;canMerge&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;GroceryLine&lt;/span&gt; &lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;GroceryLine&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;canonicalId&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;canonicalId&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="c1"&gt;// "to taste" lines have no quantity; folding them loses nothing.&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;quantity&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;quantity&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="c1"&gt;// A bare count ("2 scallions") and a counted unit ("2 stalks") are the&lt;/span&gt;
    &lt;span class="c1"&gt;// same shopping decision; a bare count and a *measured* amount&lt;/span&gt;
    &lt;span class="c1"&gt;// ("200 g eggs") are not.&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nc"&gt;UnitConverter&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;isCount&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nc"&gt;UnitConverter&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;isCount&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nc"&gt;UnitConverter&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;canConvert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;unit&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Same unit, add the quantities. A "to taste" line folds into a measured one and keeps the known amount. Different measured units merge only when they convert (volume with volume, mass with mass), and the first line's unit wins: &lt;code&gt;250 ml milk&lt;/code&gt; plus &lt;code&gt;1 cup milk&lt;/code&gt; becomes &lt;code&gt;490 ml&lt;/code&gt;. Weight and volume never merge, so &lt;code&gt;200 g chocolate&lt;/code&gt; and &lt;code&gt;1 cup chocolate&lt;/code&gt; stay two lines. Count units never convert into weights, or into each other: &lt;code&gt;2 cloves garlic&lt;/code&gt; and &lt;code&gt;1 bunch garlic&lt;/code&gt; stay two lines. A bare &lt;code&gt;3 eggs&lt;/code&gt; can merge with a count unit, but not with &lt;code&gt;200 g eggs&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it does not handle yet
&lt;/h2&gt;

&lt;p&gt;The parser does not normalize full-width digits. &lt;code&gt;２&lt;/code&gt; is not treated as &lt;code&gt;2&lt;/code&gt;, and there is no NFKC pass, so a line like &lt;code&gt;小麦粉 ２カップ&lt;/code&gt; will not get a quantity today. Japanese kanji numerals (&lt;code&gt;一&lt;/code&gt;, &lt;code&gt;二&lt;/code&gt;, &lt;code&gt;半分&lt;/code&gt;) and Sino-Korean numbers (&lt;code&gt;일&lt;/code&gt;, &lt;code&gt;이&lt;/code&gt;, &lt;code&gt;삼&lt;/code&gt;) are also not handled. Those will need work. I have not pretended they already work.&lt;/p&gt;

&lt;p&gt;The parser also does not take a language argument. A mixed or ambiguous line has only the combined tables and the "prefer no quantity" rule to fall back on. The draft you review before saving is there for the rest.&lt;/p&gt;

&lt;p&gt;Kerfuffle is free on the App Store, with 132 starter recipes and no account to create. If you want to try it on a real recipe, it is &lt;a href="https://apps.apple.com/us/app/kerfuffle-recipes/id6818037609" rel="noopener noreferrer"&gt;here&lt;/a&gt;. The site is at &lt;a href="https://hyorinmaruk.github.io/recipebox-site/" rel="noopener noreferrer"&gt;hyorinmaruk.github.io/recipebox-site&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>flutter</category>
      <category>dart</category>
      <category>i18n</category>
      <category>mobile</category>
    </item>
    <item>
      <title>I Built the Recipe App I Actually Wanted</title>
      <dc:creator>Hyorinmaru</dc:creator>
      <pubDate>Tue, 10 Nov 2026 06:02:55 +0000</pubDate>
      <link>https://dev.to/hyorinmaruk/i-built-the-recipe-app-i-actually-wanted-3b8o</link>
      <guid>https://dev.to/hyorinmaruk/i-built-the-recipe-app-i-actually-wanted-3b8o</guid>
      <description>&lt;p&gt;I cook at home most days. Rice, chicken, sometimes a pasta night when I feel like a restaurant. For a long time my recipes lived everywhere: screenshots, notes, open browser tabs, a few just in my head.&lt;/p&gt;

&lt;p&gt;Every recipe app I tried wanted something first. An account. A subscription. My recipes on their server. I just wanted my recipes on my phone, in one place, without any of that. So I built it. It's called Kerfuffle, and it's on the App Store now.&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%2F2oyasi14w2comgpitp8n.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%2F2oyasi14w2comgpitp8n.png" alt=" " width="800" height="1739"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Here's what it does:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Everything lives on your phone. No account, no subscription, no server, no analytics, no ads. Your recipes are yours.&lt;/li&gt;
&lt;li&gt;Save a recipe from the web in one tap. Share a recipe page to Kerfuffle and it arrives as a draft you review before saving. You can also import a Paprika collection, or add your own recipes with photos.&lt;/li&gt;
&lt;li&gt;Cook mode shows one step at a time in big type. If a step says simmer for 20 minutes, one tap starts a timer that keeps running even when you put the phone down.&lt;/li&gt;
&lt;li&gt;Scale the servings and the ingredients follow. Each recipe shows a rough calorie estimate too.&lt;/li&gt;
&lt;li&gt;Plan the week by dropping recipes onto days, then send the whole week to a shopping list. Ingredients from different recipes merge into one list, grouped the way a shop is laid out.&lt;/li&gt;
&lt;li&gt;It comes with 132 starter recipes, and the whole app speaks six languages: English, Spanish, Japanese, Kazakh, Korean, and Russian.&lt;/li&gt;
&lt;/ul&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%2Fjeht2mpj2mk4jjnzjlht.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%2Fjeht2mpj2mk4jjnzjlht.png" alt=" " width="800" height="1739"&gt;&lt;/a&gt;&lt;br&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%2Fa9z4bbf8bsenp9gjhnl5.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%2Fa9z4bbf8bsenp9gjhnl5.png" alt=" " width="800" height="1739"&gt;&lt;/a&gt;&lt;br&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%2Ferkg27ve0dysfln67kb4.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%2Ferkg27ve0dysfln67kb4.png" alt=" " width="800" height="1739"&gt;&lt;/a&gt;&lt;br&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%2F7e1usosncmfrpyv3qf3s.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%2F7e1usosncmfrpyv3qf3s.png" alt=" " width="800" height="1739"&gt;&lt;/a&gt;&lt;br&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%2F90163l644e7n7blmdrtv.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%2F90163l644e7n7blmdrtv.png" alt=" " width="800" height="1739"&gt;&lt;/a&gt;&lt;br&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%2Fqr66o5k5b7155ufc2v11.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%2Fqr66o5k5b7155ufc2v11.png" alt=" " width="800" height="1739"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;I built Kerfuffle for the way I actually cook. If you cook at home and want your recipes in one place without signing up for anything, try it.&lt;/p&gt;

&lt;p&gt;Get it on the App Store: &lt;a href="https://apps.apple.com/us/app/kerfuffle-recipes/id6818037609" rel="noopener noreferrer"&gt;https://apps.apple.com/us/app/kerfuffle-recipes/id6818037609&lt;/a&gt;&lt;br&gt;
Learn more: &lt;a href="https://hyorinmaruk.github.io/recipebox-site/" rel="noopener noreferrer"&gt;https://hyorinmaruk.github.io/recipebox-site/&lt;/a&gt;&lt;/p&gt;

</description>
      <category>cooking</category>
      <category>recipes</category>
      <category>ios</category>
      <category>flutter</category>
    </item>
    <item>
      <title>We Didn't Win ScienceClaw. I Still Believe in Our Project.</title>
      <dc:creator>Hyorinmaru</dc:creator>
      <pubDate>Wed, 04 Nov 2026 05:51:04 +0000</pubDate>
      <link>https://dev.to/hyorinmaruk/we-didnt-win-scienceclaw-i-still-believe-in-our-project-3ihe</link>
      <guid>https://dev.to/hyorinmaruk/we-didnt-win-scienceclaw-i-still-believe-in-our-project-3ihe</guid>
      <description>&lt;p&gt;Last weekend, my two teammates and I spent three days at the MIT Media Lab for the ScienceClaw Hackathon 2026. We slept badly, drank too much coffee, and built something I am still proud of. We did not win.&lt;br&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%2Fbrofaoav7dx6ppydw1yr.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%2Fbrofaoav7dx6ppydw1yr.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;QualityRelay at a glance: station agents, the SAFE-MESH safety gate, and the human in the loop.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Walk around the room and you would think every team was curing something. Drug discovery, enzyme design, protein folding, that kind of work. It was genuinely impressive, and it made our project look a little boring by comparison. We were in the Manufacturing and Materials track, and our project was about factory inspection. No molecules. No miracle cures. Just parts on a line, and the question of whether each one is safe to ship.&lt;/p&gt;

&lt;p&gt;Here is the problem we picked. Take a slot and a tab that must fit together. Each part can pass its own tolerance check, and the pair can still fail to fit, because the errors stack up. A caliper, a camera, and the CAD file will also happily disagree with each other about the same part. Get these calls wrong and it gets expensive fast: unplanned downtime at a large automotive plant costs around $2.3M per hour, and the cost of poor quality typically eats 15 to 20 percent of sales.&lt;br&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%2F89p63jrdl1e8h1q5ib69.jpg" 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%2F89p63jrdl1e8h1q5ib69.jpg" alt=" " width="800" height="610"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Our 3D-printed slots and tabs, labeled by design width.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Our answer is called QualityRelay SAFE-MESH, and it works the way a good engineering team works. Instead of sending every raw camera frame to one central server that decides everything, each station owns its own evidence. A caliper agent measures the critical width. A camera agent takes a photo per part. A vision agent looks for anomalies. An assembly agent pairs the slot and the tab. They exchange small signed messages, about 4.7 kB per part instead of about 4 MB of raw frames, and a safety gate turns the evidence into one of three decisions: PASS, HOLD, or REJECT. The rule we cared most about is simple. Unsure means HOLD, never PASS. Missing, stale, duplicated, or unverifiable evidence sends the part to HOLD. And a human approves every physical action. The agents recommend; people sort and sign off.&lt;br&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%2F7jmohxo1p24p6oca0cif.jpg" 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%2F7jmohxo1p24p6oca0cif.jpg" alt=" " width="800" height="391"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The digital factory with the live agent panel.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;The result we are proudest of is what happens when the network fails. We registered the test before writing the code, then ran our station agents against a central server on the same simulated line. On a fast link, both architectures performed identically, and we say so openly. But on a 2 Mbit/s uplink, the central server dropped to 43.2 good assemblies per hour while our station agents held 195.8. With the server down for 30 minutes, it was 89.7 against 195.8. Same safety in every case: 1.04 bad parts shipped per 1,000, against 3.90 for fixed rules. That is about 850 times less network traffic, and a line that keeps running when the factory network does not.&lt;br&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%2Fn5deib02gxt1cmnjw4da.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%2Fn5deib02gxt1cmnjw4da.png" alt=" " width="800" height="384"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The key result: the line keeps running when the network does not.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;We also closed the loop. In one scenario, a whole first batch comes out wrong because of a 1.6 mm tool offset error. The agents catch it, the line stops, the caliper chamber measures every suspect part, and a CNC agent computes a G-code correction that an operator approves before the machine restarts. In simulation, that run shipped 60 out of 60 good pairs with zero escapes. Feeding inspection back to a printer under drift raised good output from 154 to 195 parts per hour.&lt;/p&gt;

&lt;p&gt;None of this stayed purely virtual. We printed eight slot and tab specimens, from comfortable clearance to deliberately defective, built a caliper chamber with an ESP32 logger that streams every reading to a CSV file, ran fit tests, and collected 88 labeled photos. Our anomaly model, trained on an NVIDIA L4 GPU, scored 0.948 AUROC on the public MVTec screw benchmark. The whole thing is backed by 550 plus automated tests, and we kept our negative results in the open: our first policy was not better than conservative rules, majority voting among agents turned out to be unsafe, and one cost-saving claim got withdrawn after we found unpriced messages. Every claim in the repo is tagged as simulated, physical capture, or external dataset, so you can check what is real and what is not.&lt;br&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%2Fu9fqgqkueswrneb0hqm4.jpg" 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%2Fu9fqgqkueswrneb0hqm4.jpg" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The caliper chamber: each reading streams to a CSV row.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;So why do I still believe in it, after losing to drug discovery? Because boring problems are expensive problems. Most of the world does not work in a lab curing diseases. It works in plants where a bad part reaching a customer, or a good part held for no reason, costs real money and real jobs. A system that lets a line keep deciding, safely, when the network fails is the kind of unglamorous infrastructure that factories actually need. The MIT stage rewards the spectacular. I would rather build the thing that keeps running on a Tuesday afternoon when the uplink dies.&lt;/p&gt;

&lt;p&gt;The project is open source, and the demo runs in your browser. If you work in manufacturing, or you just like the idea of agents that say "I'm not sure, hold this part" instead of guessing, I would love to hear what you think.&lt;/p&gt;

&lt;p&gt;Repo: &lt;a href="https://github.com/Bakdaulet-omeraiuly/qualityrelay" rel="noopener noreferrer"&gt;https://github.com/Bakdaulet-omeraiuly/qualityrelay&lt;/a&gt;&lt;br&gt;
Live demo: &lt;a href="https://qualityrelay.vercel.app" rel="noopener noreferrer"&gt;https://qualityrelay.vercel.app&lt;/a&gt;&lt;/p&gt;

</description>
      <category>hackathon</category>
      <category>manufacturing</category>
      <category>robotics</category>
      <category>showdev</category>
    </item>
    <item>
      <title>I Built a Free Wongoji Editor for TOPIK Writing Practice</title>
      <dc:creator>Hyorinmaru</dc:creator>
      <pubDate>Thu, 29 Oct 2026 20:42:33 +0000</pubDate>
      <link>https://dev.to/hyorinmaruk/i-built-a-free-wongoji-editor-for-topik-writing-practice-4cjd</link>
      <guid>https://dev.to/hyorinmaruk/i-built-a-free-wongoji-editor-for-topik-writing-practice-4cjd</guid>
      <description>&lt;p&gt;Hey everyone! I was prepping for TOPIK writing and ran into a problem I couldn't stand: practicing handwriting on paper. I type everything on a keyboard, and handwriting practice felt slow and messy — exactly the wrong thing to focus on when the writing section is already hard enough.&lt;/p&gt;

&lt;p&gt;So I built the tool I wished existed: a Korean wongoji editor that takes what you type and formats it directly into authentic 원고지 (manuscript paper) — the exact grid format TOPIK uses.&lt;/p&gt;

&lt;p&gt;What it does:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Real exam practice: 21 real past TOPIK II exams, with Writing 1 (문제 53) and Writing 2 (문제 54) questions laid out on proper 원고지.&lt;/li&gt;
&lt;li&gt;Type, don't handwrite: if you think faster on a keyboard, you can focus on writing good Korean instead of fighting your handwriting.&lt;/li&gt;
&lt;/ul&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%2F5af56a6j0vaf5xrh2w0l.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%2F5af56a6j0vaf5xrh2w0l.png" alt=" " width="800" height="1035"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Type a sentence, and it lands on authentic 원고지, character by character&lt;/em&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A feedback loop that actually works: since everything is already typed, you can copy your essay out in one go and paste it into ChatGPT, Claude, Gemini, or Grok to analyze your grammar and mistakes. Fix them, rewrite, repeat — way faster than waiting for a teacher to mark up a paper essay.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It's free, and it works for TOPIK prep or just general Korean writing practice. Try it here: &lt;a href="https://korean-wongoji-editor.vercel.app/topik-prep" rel="noopener noreferrer"&gt;https://korean-wongoji-editor.vercel.app/topik-prep&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;If you try it, I'd love feedback — especially from anyone else prepping for TOPIK writing.&lt;/p&gt;

</description>
      <category>korean</category>
      <category>topik</category>
      <category>showdev</category>
      <category>webdev</category>
    </item>
  </channel>
</rss>
