MBMilos Bozicinmilosbozic.hashnode.dev路Sep 14 路 24 min readDoes an LLM Watermark Survive Editing? Testing 4 Schemes Against ParaphrasingPost 2 of 2: word edits, synonym swaps and LLM paraphrasing against a from-scratch watermark, and the trade-off between surviving edits and protecting the key. 馃摑 Notebook In the first post we built t00
MBMilos Bozicinmilosbozic.hashnode.dev路Sep 10 路 19 min readBuilding a Reliable AI Agent from Scratch with a Small Local LLM (Part 2)馃摑 Notebook In Part 1, we built a ReAct agent on Phi-4-mini (3.8B parameters) and improved it one stage at a time, measuring every stage on the same 12-question test set. The agent works in a fictiona00
MBMilos Bozicinmilosbozic.hashnode.dev路Sep 8 路 31 min readHow LLM Text Watermarking Works: Build One from Scratch in PythonPost 1 of 2: implement the green-red list watermark, detect it with a proper statistical test, and measure what it costs. 馃摑 Notebook In August 2026 Anthropic announced that new Claude models will wat00
MBMilos Bozicinmilosbozic.hashnode.dev路Sep 3 路 22 min readBuilding a Reliable AI Agent from Scratch with a Small Local LLM馃摑 Notebook Agents built on large API models make tool calling look easy. Try the same thing with a small model on your own GPU, and things fall apart: the model breaks the output format, loops until 00
