This weekend was mostly an office based weekend. Spent a lot of time on the day job, preparing for the coming week, building UI's and agent workflows. Fun, but I can't talk about the details. Sadly.
Also, mostly in parallel, I built out Blossom12. At least the first part of it.
The internal agent isn't fully online yet. But the idea here is that you have an associative memory system on the back end that works like a ticket system on the front end.
It works as an MCP, and connects to claude code or Cursor, or whatever. Runs in a happy little docker container. Tell claude what project you're working on, and to set up a ticket, or epic, and the agent has full CRU access to the system.
It can search, make plans, take notes, remember things, either granularly, using Pthia, the onboard agent to do the heavy lifting. Think of Pthia as an assistant's assistant. Part bookkeeper, part mystic oracle.
This agent is interesting because of how much performance I've managed to squeeze out of the Qwen3-0.6b model. It's absolutely remarkable, and I'm honestly a little surprised at how well it all works.
On the one hand it's a shocking amount of performance for an agent that runs in CPU and RAM. On the other... how difficult is this stuff, really?
When I'm done, there's going to be a chat in the UI, where Pthia can walk a human user through organization and ticket management. Writing tickets, updating epics, etc.
Back end for this one is written in Python, against mongo and faiss with a custom openai interface that sits on top of ollama (internal agents are fully tooled), front end is the least offensive Vite/React setup I could realistically muster.
It's stable, and it works today, but needs a couple of the more interesting features implemented before I drop it.
So... next couple weeks?