Git Visually: What I Wish I Knew Before My First Commit
From pancakes_FINAL_v3.md to rebase, fixup and bisect: one picture per idea, in the order you will actually need them.
Posts
7 articles on agentic systems, LLMs, machine learning, and software engineering.
October 2026
September 2026
Does enforcing JSON make answers worse?
Six ways to wrap the same reply, nine models, five benchmarks. What does the JSON envelope really cost your answers?
Can an LLM feel the road? Fast, cheap drive-time estimates across France
A distance curve, a neural network and an LLM all asked how long a drive takes. Which one would you trust inside an optimizer loop?
July 2026
Squashing Git Commits: Interactive Rebase, Amend, and Autosquash
A hands-on tutorial: cleaning history with interactive rebase, folding a forgotten change into the last commit with amend, and fixing an older commit with fixup and autosquash.
May 2026
Data URLs versus rotating presigned HTTPS: latency and cache in multimodal chat APIs
Benchmarks whether sending PNG bytes inline as Base64 data URLs or as Cloudflare R2 presigned HTTPS URLs changes multimodal completion latency. Three LLMs on synthetic OCR.
Cobjectric: Measuring Parsing Quality for Structured Data
A Python library for fill rate, fill-rate accuracy, and fuzzy similarity on structured payloads such as CV parses or LLM JSON.
Hello world
A short note on what I write here, and what a post is allowed to get wrong.