Sunday started with the usual quiet hum of the weekend, but the git log tells a different story. I spent most of the day deep in the trenches of crisis-management-ops, churning out fifteen commits. The rest of the stack got some love too, but the center of gravity was clearly the crisis platform. It feels good to finally have the momentum to push through the complexity.
The headline here is the multi-case architecture. I built out a case picker, a dedicated /cases page, and a dynamic case context that flows through the whole app. It was a solid day of plumbing, touching ten files to make sure the backend routes and frontend components played nice together. You can switch contexts now, which is a huge win for usability. I also spent some time on the knowledge graph, implementing a 3D force-directed view. It’s not just a chart anymore; it’s interactive and actually helps visualize the connections in the data.
Another major arc was the evidence pipeline. I added a drag-and-drop upload zone with status badges, so users can see exactly where their files are in the process. Behind the scenes, I built an OCR sidecar using ocrmypdf to extract text from PDFs, and an in-process worker that handles per-page chunking. It’s a bit of a beast under the hood, but it means the system can now ingest and index evidence much faster. I also wired up citations, so when the AI references a source, it’s a clickable link that takes you right to that page in the PDF.
I built out a case picker, a dedicated /cases page, and a dynamic case context that flows through the whole app.
I didn’t forget about the drafts system. It now supports DOCX exports alongside PDFs, powered by a pandoc subprocess in the brigade-pdf-gen repo. I added a download button for DOCX on the drafts page and persisted the files alongside the PDFs. It’s a small feature, but it makes a big difference for people who need to edit the output. I also added a model picker for per-call drafting and some search filters on the drafts page.
The music stack got a fresh coat of paint too. I pushed v0.4.0 to majiks.music, which shifts to an anonymous-first model. You don’t need to sign in to browse or listen anymore; auth is only required for publishing and settings. I also shipped v0.3.0, which brings multi-variant generation, advanced parameters, ratings, and a lyrics editor. The player transport bar is finally here, with play, pause, scrubbing, and volume controls. It feels like a real product now, not just a prototype.
I also touched majiks-studio-web to add a multipart publish endpoint with audio support and a Bridge tier check. On majiks.online, I added a subscription route for the studio bridge. And in unicorn-brigade, I released versions 1.8.0, 1.9.0, and 1.10.0, which bring per-call model overrides, mission ID correlation, and per-case memory scoping. It’s a lot of releases, but they’re all necessary to keep the agents behaving themselves.
Also today: I fixed a missing folder icon in the layout and added an empty-submit validator to the chat panel. Small things, but they keep the UI from looking broken.
It was a long day, but the platform feels more cohesive. The case management, evidence ingestion, and drafting tools are finally clicking into place. The music app is polished and ready for wider use. I’m happy with what shipped.
Real product captures — click any to enlarge.