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.
Three Ways I Tried to Automate 1Point3Acres Daily Check-In and Quiz Answering
Direct Codex browser control, a supervised WebContentAgent library, and a fully local LLM workflow—what failed, what I use now, and what I am still exploring.
[2/2] How I Made a Codex-to-ChatGPT Bridge Verifiable
A local bridge sends one authorized request and brings back a checked answer. Its failures explain what message identity, durable state and realistic tests must get right.
[1/2] (Deprecated) How I Let Codex Talk to ChatGPT Through My Signed-In Chrome
The first prototype made the round trip work. Its failures taught me why browser mechanics need a smaller, testable home.
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
Experiments that turn Olden Era Arena screenshots into structured game state, grouped by the evidence they use and shown in conceptual order.
ApproachVisible text · ordered workflow
OCR Scene Understanding
Read visible text to identify the current screen, then keep workflow progress separate from current-frame evidence.
- Text and position
- Current screen
- Phase context
Understanding Game Screens with OCR and a Small State Machine
Text and position identify the current screen; a separate temporal model tracks progress. A two-video prototype shows why this is both a perception and a software-systems problem.
Recognizing Game Scenes and Choosing Review Frames
A design record for classifying Arena screens, handling incomplete pages, and reducing a long video to meaningful UI changes.
ApproachStable UI regions · insufficient text
Visual Scene Fallback
Recognize loading screens and battlefields from combinations of UI regions, while keeping content readiness and uncertainty explicit.
- OCR uncertainty
- Visual evidence
- Scene or abstention
ApproachKnown geometry · nearly fixed camera
Calibrated Grid Registration
Use the battlefield’s regular hexagonal geometry and a saved calibration to infer cell positions, then refine their alignment from screenshot edges.
- Calibrated grid
- Image edge scoring
- Small alignment correction
ApproachKnown position
Fixed-Crop Classification & Rejection
Assume each target icon appears at a known UI location. With localization removed, perception becomes a classification-and-rejection problem.
- Known UI slot
- Fixed crop
- Classify + reject
ApproachUnknown position · learned localization
Learned Object Detection
Do not assume the icon location is known. A learned detector proposes boxes; a separate catalog matcher suggests identities. The current evidence is a small selection-screen pilot.
- Full screenshot
- Learned detector
- Boxes + identities
ApproachKnown layout · readable titles
Printed-Name Recognition
Read the name printed beside an icon, resolve bilingual aliases, and propose an identity for human review. Readable text does not prove crop validity.
- Title region
- Local OCR
- Canonical name + review
Reading Arena Choice Names with OCR Instead of Guessing the Artwork
Fixed title regions and bilingual aliases turn printed game names into review suggestions. Text matching helps annotation, but does not prove identity accuracy or image validity.
Choosing a Fast, Accurate OCR Pipeline for Small Game UI Titles
An exploratory failure study compares four OCR paths on eleven real game crops, separates transcription from input validity, and measures startup and resident end-to-end speed.
ApproachUnknown position · code-generated candidates
Sliding-Window Icon Detection
Search the full screenshot with many candidate crops, encode each crop into visual features, and match them against the known icon catalog.
- Full screenshot
- Candidate crops
- Catalog matching
Sliding-Window Icon Detection with a Catalog-Trained Backbone
Keep code-generated search windows, train the visual backbone on 390 catalog images, and learn crop quality separately. Gameplay screenshots do not update model weights.
Sliding-Window Icon Detection with a Public Vision Backbone
Search a full screenshot with code-generated candidate crops and a public pretrained visual model. No game-specific training; useful identity matches, but unreliable crop alignment.
Game Probability — Olden Era
Reviewed game tables turned into transparent Arena odds, with the event and evidence boundaries kept explicit.