Safouane Chergui
AI Engineer. I write about ML, LLM systems, evals, and Python tooling.
Weekends Eureka
Posts tagged · show all
-
Things I wish I'd known about uv
A year of uv lessons: uv add vs uv pip install, the default dev group, locked syncs in CI, and run, tool and uvx.
-
Specs First, Prompts Second
LLM evals lecture 1: the three gulfs, and why the spec comes before the prompt.
-
Why filtering breaks your vector search
Why a WHERE filter can break HNSW search, and how Qdrant's filterable HNSW fixes it.
-
Why does SwiGLU work ?
Why modern LLMs use SwiGLU instead of ReLU in their feed-forward layers, one question at a time.
-
Stop Vibe-Checking: Real-World Lessons on LLM Evals
Lessons from deploying LLM systems in production on evaluating them beyond vibe checks.
-
PDF parsing is hard
What a PDF is internally, why parsing one is hard, and how Docling handled a real project.
-
Why I ditched pip and conda for Pixi
Why I moved from pip and conda to pixi: features, environments, tasks and configuration.
-
Byte Pair Encoding Tokenization
How byte pair encoding works, by building a basic byte-level tokenizer from scratch.
No matching items