What I’m building on my own time
I build things outside my client work to stay current with what these tools can actually do. Three of them from the past few months seem worth passing along, because each one suggests something a research organization might do with the same tool. All three are running today in my practice.
I keep everything I know about my work (projects, people, commitments, meeting notes) as a set of files an AI reads and writes. I call the system Jarvis. A Raspberry Pi, a small always-on computer, makes Jarvis reachable from my phone: from anywhere with cell coverage I can ask what came out of last week’s meetings, capture a half-formed idea, or have it draft something, and whatever it learns is waiting at my desk the next morning.
A research team’s institutional memory could work the same way. Interview notes, decisions, drafts, and the literature a team has digested could become something a researcher can question from the field, on the phone already in their pocket.
Local models, the ones that run entirely on your own machine, have become quite good over the past few months. The work happens on the machine itself, in real time. Nothing touches an external server, including the AI company’s.
Research buildings hold plenty of data that can’t leave the room, and I keep client material with the same constraint. A model that runs with no connection to the outside serves exactly that case: a team could transcribe and query sensitive interviews with no cloud involved, or work a restricted dataset knowing nothing can get out. A well-equipped laptop is enough to run it.
An AI can now drive a real browser: the same Chrome window a person uses, with that person’s own logins. The tool is Anthropic’s Claude in Chrome, and my guess is that of the three, this is the one people are least likely to know exists. I recently had it read three months of my own work, draw up a revised narrative for my website from what it read, and then go into Squarespace (the website editor) and make the edits itself, so the analysis and the follow-through happened in one motion.
Last week I handed it a chore: booking a summer’s worth of conference rooms through my coworking space’s reservation site. It wrote a plan and asked me to approve it. Then it worked the calendar and booking forms on its own, and checked back before spending my money.
Most of the systems a researcher touches, the web portals, records systems, and submission sites, were built for a person with a mouse, with little other way in. A tool that can operate a browser could work almost any of them.
