Skip to main content
Diving deep into this “deploy it myself and see how much I can run before the computer breaks” situation. Really bummed I don’t have the compute power to locally deploy models, but it’s too expensive at the current time. Still looking for the right solution to have a pool of agents that act autonomously. There’s a combination of pieces that all need to fit together to be a good solution, and no orchestrator or library I’ve seen so far can do it. Off the top of my head, it involves: intelligent, dynamic invocation, separation of work performed and communication with the user, A2A collaboration, memory and learning, control panel for creating and updating agents, remotely hosted, observability through a proxy, individual permissions and resource access, and so forth. Back to the main topic. What have I deployed in April? The biggest change was that I migrated everything to docker swarm. Before, I had docker compose files on each machine. The entire folder with all compose files was duplicated to both machines, but only some of the containers were running, depending on the machine, tracked in the docs. Not ideal. With swarm, I’ve got now got a 2 node cluster. Need to diagram this for how services are spun up. Best scenario would be an agent that dynamically distributes the services to specific machines ad hoc (all media streaming goes on i3 since pentium has transcoding issues). The other thing I did to make media streaming more robust was to implement mergerfs to create a storage pool for the i3. I added a chron job on top of that to continually check if my drives are healthy, and I put that in Uptime Kuma and set it up with Telegram notifications, so now I get pinged whenever a service or drive goes down (very handy, since drives are going out like once a month). All right, now onto the cool stuff. N8N! Was excited about this one for a while. Fits into the sequential AI workflows category. Quite a few ideas on other ones to build, but the first was an AI classifier for my emails. If you’ve been following along, I used to be inbox-zero, and I’ll usally get there every few weeks or so, keep it for a week, and repeat. However, I have a deep and burning hatred for emails that I didn’t ask for. As such, the first priority of my classifier is to ask two questions: is there an unsubscribe button? and, would a regular person delete this email without reading it? if yes to either, it goes into a respective folder, and then there’s additional folders for receipts, travel, orders, etc. I still have my own email server, runnning on Stalwart on my local machine. Gotta say, I was nervous having something as important as email running on an old computer in my living room, but it’s been working spectacularly. Other fun ones. I’ve had a working version of excalideck (might change the name since there’s another repo with ~200 GH stars), a self hosted excalidraw diagram manager, for some time now. I did a pretty big revamp on it, and it’s pretty much ready to be publicized, with a few changes. I need to seek help on how to get it broadcasted though, since I have no idea how to do that and posting on HN and reddit threads are full of junk and AI posts now that I’m not sure what the right way to do it is. Perhaps just commenting on reddit threads where people ask about Excalidraw in general. Anywho, what else have we got? Bulwark. Very clearly a vibe-coded app, but it’s not bad. It surpasses Twake and Jmap-webmail in features, so I use Bulwark. I definitely think it could have a prettier UI, but it works for now. I did make one change to the source code, I added in the ability to respond from the email address that received the message, useful since I create new emails for every online account, all under @augustin.ai. I had trouble in the past responding to support emails sent from a different address, and the case would be lost. Got some positive feedback on that PR.