Feiyu's Technical Blog

Feiyu's Technical Blog

Notes on software, systems, and things I am learning.

Agents & Automation

Local models, tool-using agents, and practical information-extraction workflows.

  1. Automating 1Point3Acres Daily Check-In and Quiz Answering with a Local LLM

    A Windows app runs around 9:00 AM, uses signed-in Chrome to complete a forum check-in, and starts a local LLM only when a quiz needs an answer.

  2. [2/2] How I Made a Codex-to-ChatGPT Bridge Verifiable

    A narrow local bridge reuses my signed-in ChatGPT conversation, records the irreversible send boundary, and rejects unverified extraction instead of guessing.

  3. [1/2] (Deprecated) How I Let Codex Talk to ChatGPT Through My Signed-In Chrome

    My first Codex-to-ChatGPT bridge let Codex control Chrome directly. It proved the idea worked, but the design was too fragile to keep.

  4. Local LLM + Agent for Web Information Extraction

    A local agent architecture for semantic web extraction, browser automation, bounded model context, and durable task state.

Game Perception — Olden Era

Computer-vision experiments that turn arena gameplay screenshots into structured game state.

  1. Do Verified Gameplay Crops Improve a Catalog-Trained Game UI Icon Classifier?

    A controlled one-seed development experiment comparing catalog-only training with verified gameplay supervision, and a learned validity gate with classification-confidence thresholding.

  2. 从一次模型退化审计到可复现实验：重构游戏 UI 图标分类器

    我们冻结 2,136 个可追溯 crop，用五个随机种子比较 catalog-only、gameplay adaptation 与 learned rejection，并逐样本审计 B04 的结果和边界。

  3. Training a Game UI Icon Detector Using Only the Icon Catalog

    Using 390 catalog images—and no gameplay screenshots for weight updates—we test full-screen hero and skill detection.

  4. Full-Image Icon Search with Sliding Windows and Visual Embeddings

    In this seven-frame experiment, the search often identifies the right hero or skill, but its highest-scoring box is not always the best crop.