<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Engineering on Subculture Science Research Group&#39;s Blog</title>
    <link>https://www.subcul-science.com/en/tags/engineering/</link>
    <description>Recent content in Engineering on Subculture Science Research Group&#39;s Blog</description>
    <generator>Hugo</generator>
    <language>en-US</language>
    <lastBuildDate>Sun, 15 Nov 2026 15:20:00 +0900</lastBuildDate>
    <atom:link href="https://www.subcul-science.com/en/tags/engineering/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Co-evolving Four Roles, GAN-Style: The Testbed Behind umigame_GAN</title>
      <link>https://www.subcul-science.com/en/post/20261115umigame-gan-tech/</link>
      <pubDate>Sun, 15 Nov 2026 15:20:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115umigame-gan-tech/</guid>
      <description>A testbed for strengthening trial-and-error reasoning in small LLMs: four competing roles, a domain-independent loop, R-Zero and EDA style rewards, and an MLX LoRA comparison design. Code excerpts and measured numbers from the repository.</description>
    </item>
    <item>
      <title>Detecting Duplicate Images with dHash and Hamming Distance in a Chrome Extension</title>
      <link>https://www.subcul-science.com/en/post/20261115x-original-filter-tech/</link>
      <pubDate>Sun, 15 Nov 2026 15:00:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115x-original-filter-tech/</guid>
      <description>How the X Original Filter Chrome extension hides reposts and duplicate images on X: a 64-bit dHash built from a 9x8 thumbnail, Hamming distance matching, an IndexedDB record of first sightings, and a rule that the earlier timestamp wins.</description>
    </item>
    <item>
      <title>Loading 80 Parcels into a Kei Van in Seconds: Extreme Points and an Independent Validator</title>
      <link>https://www.subcul-science.com/en/post/202611153d-packing-tech/</link>
      <pubDate>Sun, 15 Nov 2026 14:40:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/202611153d-packing-tech/</guid>
      <description>How a 3D load-planning demo builds stacks that respect delivery order, this-side-up, load bearing, and support ratio. Extreme points plus GRASP generate candidates, and a validator written independently of the solver decides whether a plan is feasible.</description>
    </item>
    <item>
      <title>Finding the Original of a Reposted Image: How Origin Finder Combines Five Reverse Image Searches</title>
      <link>https://www.subcul-science.com/en/post/20261115origin-finder-tech/</link>
      <pubDate>Sun, 15 Nov 2026 14:20:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115origin-finder-tech/</guid>
      <description>How Origin Finder calls five reverse image search engines in parallel, decides which candidates are the same image in staged passes, and returns the oldest post as the origin. Code excerpts and measured numbers from the repository.</description>
    </item>
    <item>
      <title>Turning a Drawing-Only Part Back into Editable 3D: How image2CAD Keeps Dimensions and History</title>
      <link>https://www.subcul-science.com/en/post/20261115image2cad-solution/</link>
      <pubDate>Sun, 15 Nov 2026 14:10:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115image2cad-solution/</guid>
      <description>When the CAD file is gone and only a drawing survives, modelling the shape by hand is not enough: the first dimension edit breaks it. How image2CAD treats drawing dimensions as the source of truth, tests editability mechanically, and uses three independent checks to avoid false accepts.</description>
    </item>
    <item>
      <title>From 2D Drawings to Editable FreeCAD Models: the image2CAD Pipeline</title>
      <link>https://www.subcul-science.com/en/post/20261115image2cad-tech/</link>
      <pubDate>Sun, 15 Nov 2026 14:00:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115image2cad-tech/</guid>
      <description>image2CAD turns a drawing PDF or scan into a history-based FreeCAD PartDesign model. Two Python processes talk only through a JSON IR, dimensions stay as expressions, and three checks decide which candidate gets accepted. A tour of the design with code from the repository.</description>
    </item>
    <item>
      <title>Building a Talking 3D Character App with FastAPI, TypeScript, VRM and Passkeys</title>
      <link>https://www.subcul-science.com/en/post/202611153d-chatting-tech/</link>
      <pubDate>Sun, 15 Nov 2026 13:40:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/202611153d-chatting-tech/</guid>
      <description>How 3D Chatting is put together: FastAPI middleware ordering and passkey registration, VRM handling in TypeScript and three.js, and an MCP server that operates VRoid Studio through screenshots and OCR.</description>
    </item>
    <item>
      <title>Trend Analysis with Next.js and the YouTube Data API — Designing Filters and Aggregation</title>
      <link>https://www.subcul-science.com/en/post/20261115youtube-analyser-tech/</link>
      <pubDate>Sun, 15 Nov 2026 13:20:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115youtube-analyser-tech/</guid>
      <description>How a Next.js app filters YouTube trend videos and aggregates views, engagement rate, and a viral coefficient: a Route Handler that keeps the API key server-side, a three-call fetch chain, and a quota estimate the README undercounts.</description>
    </item>
    <item>
      <title>A Travel Calendar in Go, Gorilla Mux, and React: How BookingButler Is Built</title>
      <link>https://www.subcul-science.com/en/post/20261115booking-butler-tech/</link>
      <pubDate>Sun, 15 Nov 2026 13:00:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115booking-butler-tech/</guid>
      <description>How BookingButler turns Google Calendar events into departure times: a Go + Gorilla Mux API that exchanges the auth code, resolves stations with Google Maps, and fetches routes from NAVITIME. The design, the code, and the gaps left in it.</description>
    </item>
    <item>
      <title>Inside chatGPT_streamlit: A Search-Enabled Chat UI with Streamlit and LangChain ReAct</title>
      <link>https://www.subcul-science.com/en/post/20261115chatgpt-streamlit-tech/</link>
      <pubDate>Sun, 15 Nov 2026 12:40:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115chatgpt-streamlit-tech/</guid>
      <description>A tour of chatGPT_streamlit: a LangChain ReAct agent with Google and Wikipedia tools behind a Streamlit chat UI. How the tools, prefix/suffix prompt, and session_state history are wired, plus the version pinning and the gaps between README and code.</description>
    </item>
    <item>
      <title>Building the TopicDigestX Backend on FastAPI and PostgreSQL</title>
      <link>https://www.subcul-science.com/en/post/20261115topic-digestx-tech/</link>
      <pubDate>Sun, 15 Nov 2026 12:20:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115topic-digestx-tech/</guid>
      <description>How TopicDigestX wires FastAPI, PostgreSQL, and a planned Alembic layer: Pydantic Settings, one SQLAlchemy model, and the dependency gaps that still block the database path.</description>
    </item>
    <item>
      <title>Reversing the Order of House Hunting: Designing a Gym-First Search</title>
      <link>https://www.subcul-science.com/en/post/20261115muscle-realestate-solution/</link>
      <pubDate>Sun, 15 Nov 2026 12:10:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115muscle-realestate-solution/</guid>
      <description>House hunting starts with the office; the gym comes after the move. Muscle Real Estate flips that order into four stages: facilities, commute area, gym, property. The design, its trade-offs, and what is still mock.</description>
    </item>
    <item>
      <title>Choosing a Home From the Gym Outward: A Go (Gin) &#43; React API for Commute-Aware Search</title>
      <link>https://www.subcul-science.com/en/post/20261115muscle-realestate-tech/</link>
      <pubDate>Sun, 15 Nov 2026 12:00:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115muscle-realestate-tech/</guid>
      <description>Muscle Real Estate reverses the usual house hunt: pick a gym first, filter by commute area, then look at nearby properties. The layering in Go and Gin, the PostGIS complex search, and the commute API that is still mock.</description>
    </item>
    <item>
      <title>A LoRA Training Pipeline for Japanese LLMs on MLX: LFM2.5-1.2B-JP</title>
      <link>https://www.subcul-science.com/en/post/20261115ft-llm-tech/</link>
      <pubDate>Sun, 15 Nov 2026 11:40:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115ft-llm-tech/</guid>
      <description>How FT_LLM fine-tunes LFM2.5-1.2B-JP with LoRA on Apple Silicon: config resolution, two execution modes, a chunked loss for long sequences, early stopping, and character-level evaluation, with code excerpts and measured numbers.</description>
    </item>
    <item>
      <title>Optimizing a Band Setlist with OR-Tools TSP: Cost Matrices and Break Insertion</title>
      <link>https://www.subcul-science.com/en/post/20261115setlist-automation-tech/</link>
      <pubDate>Sun, 15 Nov 2026 11:20:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115setlist-automation-tech/</guid>
      <description>How SetlistAutomation turns song ordering into a traveling salesman problem: a transition cost matrix, an OR-Tools search configuration, YouTube Data API durations, and a break-inserter that hides the heaviest instrument swap.</description>
    </item>
    <item>
      <title>A Japanese Voice-Clone TTS Pipeline with TADA, FastAPI, Whisper, and FFmpeg</title>
      <link>https://www.subcul-science.com/en/post/20261115tts-tech/</link>
      <pubDate>Sun, 15 Nov 2026 11:00:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115tts-tech/</guid>
      <description>How a reference-audio voice-clone TTS is built with TADA 3B-ML, FastAPI, Whisper, and FFmpeg: registration in a CLI, generation in a server, and why prompt.pt is passed between them.</description>
    </item>
    <item>
      <title>A Discord Bot That Lives in the Chat: Ollama, discord.js, and a News Pipeline</title>
      <link>https://www.subcul-science.com/en/post/20261115discord-bot-tech/</link>
      <pubDate>Sun, 15 Nov 2026 10:40:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115discord-bot-tech/</guid>
      <description>How Lobster, a Kansai-dialect Discord bot, is built: one Node process running discord.js and Ollama, persona composition, an instant 👀 acknowledgement, 17 RSS feeds, and interests that change from chat.</description>
    </item>
    <item>
      <title>A Drawing Critique App in Go, SQLite (WAL), React, and an Ollama VLM</title>
      <link>https://www.subcul-science.com/en/post/20261115dessan-sensei-tech/</link>
      <pubDate>Sun, 15 Nov 2026 10:20:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115dessan-sensei-tech/</guid>
      <description>Dessan-sensei scores a drawing photo on six axes out of ten using a local vision model. A tour of the four layers — Go, SQLite in WAL mode, React, and Ollama — through the prompt builder, the JSON parser, and the transactional save.</description>
    </item>
    <item>
      <title>How JamSlicer Splits Long Videos with FFmpeg silencedetect</title>
      <link>https://www.subcul-science.com/en/post/20261115jam-slicer-tech/</link>
      <pubDate>Sun, 15 Nov 2026 10:00:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115jam-slicer-tech/</guid>
      <description>A technical tour of JamSlicer: one FFmpeg pass for volume analysis, segment calculation in the browser, and stream-copy cutting. Why each layer is split the way it is, with code from the repository.</description>
    </item>
    <item>
      <title>Building a LINE Butler Bot with Gemini Function Calling and MCP</title>
      <link>https://www.subcul-science.com/en/post/20261115turtle-buttler-tech/</link>
      <pubDate>Sun, 15 Nov 2026 09:40:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115turtle-buttler-tech/</guid>
      <description>How turtle-buttler, a LINE bot that speaks as a Kansai-dialect butler, is put together: Gemini 2.0 Flash function calling, MCP servers backed by Firebase Realtime Database, and an HTTP fallback for Cloud Run.</description>
    </item>
    <item>
      <title>Rediscovering Likes with Similar Images: Next.js, pgvector, DINOv2, and an MV3 Extension</title>
      <link>https://www.subcul-science.com/en/post/20261115x-fav-gallery-tech/</link>
      <pubDate>Sun, 15 Nov 2026 09:20:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115x-fav-gallery-tech/</guid>
      <description>How Refav Gallery turns X likes into a searchable image gallery: passive interception in a Manifest V3 extension, DINOv2 embeddings on Modal, a pgvector HNSW/exact-KNN split, and MMR for diversity, with code and measured numbers from the repository.</description>
    </item>
    <item>
      <title>How tsuzukan Tracks Manga Sequels: Go, React 19, and Three Bibliographic APIs</title>
      <link>https://www.subcul-science.com/en/post/20261115tsuzukan-tech/</link>
      <pubDate>Sun, 15 Nov 2026 09:00:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20261115tsuzukan-tech/</guid>
      <description>A tour of tsuzukan&amp;rsquo;s internals: a Go 1.26 + Gin API, a React 19 frontend, PostgreSQL, Firebase Auth, and three bibliographic sources with very different personalities — openBD, NDL Search, and Rakuten Books — folded behind one interface.</description>
    </item>
    <item>
      <title>On the use of distributed representations bagging for class classification and generalization performance</title>
      <link>https://www.subcul-science.com/en/post/20210204classifier/</link>
      <pubDate>Thu, 04 Feb 2021 13:24:19 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20210204classifier/</guid>
      <description>&lt;p&gt;After the distributed representation has been obtained, the&#xA;After the distributed representation is obtained, machine learning can be used to classify it.&lt;/p&gt;&#xA;&lt;p&gt;Models that can be used include&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Decision Tree&lt;/li&gt;&#xA;&lt;li&gt;SVM Support Vector Machine&lt;/li&gt;&#xA;&lt;li&gt;NN Neural Networks&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;and others.&lt;/p&gt;&#xA;&lt;p&gt;SVM is included in NN in a broad sense.&lt;/p&gt;&#xA;&lt;p&gt;In this section, we will use the decision tree method.&lt;/p&gt;&#xA;&lt;h2 id=&#34;bagging&#34;&gt;Bagging&lt;/h2&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Image of majority voting with multiple decision trees&lt;/li&gt;&#xA;&lt;li&gt;Simple theory&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Decision trees are highly explainable and are a classic machine learning model.&lt;/li&gt;&#xA;&lt;li&gt;Computational load is light compared to deep learning&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Depends on the size of the model&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;/li&gt;&#xA;&lt;li&gt;Not much explainability&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Do we want to analyze each of the multiple decision trees?&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;``py&#xA;from sklearn.ensemble import BaggingClassifier&#xA;from sklearn.tree import DecisionTreeClassifier&lt;/p&gt;</description>
    </item>
    <item>
      <title>Using BART (sentence summary model) with hugging face</title>
      <link>https://www.subcul-science.com/en/post/20210119huggingface/</link>
      <pubDate>Tue, 19 Jan 2021 03:03:11 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20210119huggingface/</guid>
      <description>&lt;ul&gt;&#xA;&lt;li&gt;BART is a model for document summarization&lt;/li&gt;&#xA;&lt;li&gt;Derived from the same transformer as BERT&lt;/li&gt;&#xA;&lt;li&gt;Unlike BERT, it has an encoder-decoder structure&#xA;&lt;ul&gt;&#xA;&lt;li&gt;This is because it is intended for sentence generation&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;This page shows the steps to run a tutorial on BART.&lt;/p&gt;&#xA;&lt;h2 id=&#34;procedure&#34;&gt;Procedure&lt;/h2&gt;&#xA;&lt;ol&gt;&#xA;&lt;li&gt;install transformers&lt;/li&gt;&#xA;&lt;/ol&gt;&#xA;&lt;p&gt;Run ``sh&#xA;pip install transformers&lt;/p&gt;&#xA;&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; class=&#34;chroma&#34;&gt;&lt;code class=&#34;language-fallback&#34; data-lang=&#34;fallback&#34;&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;Run summary&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;2. Run the summary&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;from transformers import BartTokenizer, BartForConditionalGeneration, BartConfig&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;model = BartForConditionalGeneration.from_pretrained(&amp;#39;facebook/bart-large&amp;#39;)&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;tokenizer = BartTokenizer.from_pretrained(&amp;#39;facebook/bart-large&amp;#39;)&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;ARTICLE_TO_SUMMARIZE = &amp;#34;My friends are cool but they eat too many carbs.&amp;#34;&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;inputs = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors=&amp;#39;pt&amp;#39;)&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;# Generate Summary&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;summary_ids = model.generate(inputs[&amp;#39;input_ids&amp;#39;], num_beams=4, max_length=5, early_stopping=True)&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids])&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;```&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;On 2021/01/18, the output was MyMy friends.&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;Interesting.&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;## Where I got stuck.&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;Error when the version of pytorch is different from the one specified in transformers.&#xA;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;pip install -U torch&lt;/p&gt;</description>
    </item>
    <item>
      <title>I made a summary text generation AI for making short-form news</title>
      <link>https://www.subcul-science.com/en/post/20200613abstractnews/</link>
      <pubDate>Sat, 13 Jun 2020 00:24:00 +0900</pubDate>
      <guid>https://www.subcul-science.com/en/post/20200613abstractnews/</guid>
      <description>&lt;p&gt;.&lt;/p&gt;&#xA;&lt;p&gt;We have successfully trained a model to automatically generate titles from news texts using a machine translation model based on deep learning.&lt;/p&gt;&#xA;&lt;h2 id=&#34;preliminaries&#34;&gt;Preliminaries&lt;/h2&gt;&#xA;&lt;p&gt;In the past, I was involved in a project to automatically generate titles from manuscripts for online news. In the past, I was involved in a project to automatically generate titles from manuscripts for online news.&lt;/p&gt;&#xA;&lt;p&gt;In order to tackle this project, I was looking into existing methods.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
