F“Francesco “oha” Rivettiinblog.oha.it·Jul 8, 2025 · 5 min readGrammarsSo far in this series, we’ve talked about LLMs, intermediate scripts, and sandboxing execution. When I was designing the system I knew LLMs were unreliable at generating code. Moreover, I was not sure how the binding of data would have worked. I was ...00
F“Francesco “oha” Rivettiinblog.oha.it·Jun 28, 2025 · 2 min read"What Was I Thinking?"read the first part Two years ago, answering questions wasn’t enough. I wanted an AI that could do things. But I didn’t trust it to do those things directly. Instead of treating the LLM as the system's operator, I used it as a translator: turning hum...00
F“Francesco “oha” Rivettiinblog.oha.it·Jun 28, 2025 · 2 min readI Wrote This in 2023. Nobody Cared. Now Everyone's Doing It.Two years ago, I started exploring how LLMs could move from text completion to real-world orchestration. Actually doing things, not just talking. Way before it was trending. I built a prototype. In Go. With a custom JavaScript-like scripting language...00
F“Francesco “oha” Rivettiinblog.oha.it·Jun 27, 2025 · 4 min readMemory-Efficient PyArrow DatasetsAfter spending months figuring out how to handle datasets in a memory-efficient way, I decided to collect everything in one place. Hopefully, this will save someone else time. What Is a Dataset? If you work with Parquet files, you know each file has...00
F“Francesco “oha” Rivettiinblog.oha.it·Jun 24, 2025 · 2 min readb-Square: Efficient Geospatial Indexing for Tabular DataAt HUB Ocean, we work with massive volumes of geospatial marine data, with some datasets containing billions of rows. To ensure fast queries, we developed b-Square — a simple but effective mechanism to index geometries in tabular formats like Parquet...00