This is my home site.
During the last IETF in Shenzhen, on a Saturday after the meeting, I went to Dafen, an artist village where painters produce handmade copies of well known oil paintings, specially Van Gogh and the usual classics. I stood watching a couple of them for a while. They are incredibly skilled and the copies are beautiful. Of course to produce so many they take shortcuts, also because they have painted…
I wanted to talk to a corpus of YouTube videos the way I talk to my Obsidian vault. One author, a few hundred videos, questions like “what do you think about housing?” that chunk-based RAG is never going to answer well. That became tertulia.jaime.win: 202 videos in, 200 structured notes and a topic graph out, served as a chat from a single Cloudflare Worker.
I built a terminal email client for Fastmail. It started as a stateless CLI for scripting and agent workflows, then grew a TUI with vim keys and an AI chat pane. It talks JMAP (RFC 8620, RFC 8621) directly.
Agents will need to work with human conversation data. Emails, call transcripts, chat logs. The naive approach is to dump everything into the LLM’s context window and hope for the best. That doesn’t scale, and it’s expensive. Better to build a knowledge base first, then let the agent query it.
Over the past few weeks I built several agent systems. I kept notes on what worked, what failed, and what made them practical to use day to day.
This Christmas I built cuentos.jaime.win from my phone. My daughter was learning to read and I wanted something interactive for her practice. The entire stack was just my iPhone, Claude Code Premium, Midjourney for some images, Vercel, and a CNAME subdomain. Built it between family dinners. I also made mira.jaime.win, a food scanner I use when grocery shopping in Finnish - convenient for finding…
A quiet change that AI agents are bringing is how we think about user interfaces. The traditiona old concept of web mashups might make a comeback with AI Agents. When agents interact with APIs dynamically (with no a priori knowledge of the subset of APIs they will use), something more fluid emerges: interfaces that generate themselves based on what the agent discovers and decides to do! These…
I presented at IETF 123 T2TRG in Madrid on agentic AI operation of IoT systems. The talk showed a working system where an AI agent discovers, reads, writes, and subscribes to IoT devices over CoAP with zero prior knowledge of the deployment. No configuration files. No device manifests. No MCP bridge. Just an entry point and standardized affordances.
Like many researchers, I’ve struggled with information overload. Between work projects, technical deep-dives, and random ideas worth remembering, I needed a system that could handle it all without becoming a burden.
Thread networking is quietly becoming ubiquitous. Every iPhone 16 and newer iPad already has Thread support built-in, Android smart hubs are shipping with it, and home routers increasingly include it by default. Given this momentum, I wanted to understand how Thread actually works under the hood—specifically, how it uses CoAP.