Posts
All the articles I've posted.
-
NDT Defect Classification Hands-on — From PyTorch + ResNet18 Transfer Learning to Grad-CAM
An opportunity came up to put my embedded object recognition experience to use, so I picked up NDT (non-destructive testing) again. A hands-on record of training a 6-class defect classifier on the NEU Steel Defect dataset using transfer learning, and visualizing the model's decision basis with Grad-CAM.
-
Claude Vision — How Do You Send Images In, and Which Model Should Receive Them
I tried both ways of sending images — base64 encoding and URL — and found them accurate but more token-costly, hit limits with small text and complex shapes, and got a sense of when to move up to a bigger model for tasks that need accuracy.
-
Giving LLMs Tools — Claude Tool Use and the Agent Loop
LLMs don't know large-number arithmetic or today's weather. I tried out Tool Use, which patches that weakness with external function calls. Message flow, automatic multi-tool selection, and the agent loop.
-
Managing AI Sessions in Vibe Coding — How to Split Common vs. Specialized Instructions
A problem I ran into managing multiple Claude Code sessions like company departments: instruction management. I worked out how much should live in the common CLAUDE.md, and where the line is for splitting off specialized CLAUDE.md files in each working folder.