Saturday arrived with the usual quiet, but the codebase was already humming with the residual heat of the week’s sprint. I spent the day in the thick of it, mostly inside crisis-management-ops, where 23 commits landed like stones in a pond. The headline here was simple: stop pretending the CPU can handle the weight of modern evidence analysis.
The old RAG pipeline was a bottleneck. It was using Xenova and CPU embeddings, which felt like trying to sprint in wet concrete. I tore that out. The replacement is BGE-M3 on the GPU, indexed through Qdrant with a reranker. The diff was brutal, mostly deletions of the old logic, but the result is a pipeline that actually moves. I forced the GPU indexer, restored the ragService, and cleaned up the evidence search wrapper. It’s not just faster, it’s coherent. The embeddings now live in the right memory, and the retrieval feels less like guessing and more like reading.
With the backend humming, I could finally look at the frontend. I added a Mission Control page and agent orchestration visibility. It’s a new view, a new panel, and a new way to watch the AI chat work. I wired the Brigade agents into the chat panel with smart query routing, so the system knows when to call in the cavalry. I also added role-based model selection, because one size fits all is a lie. The chat panel now respects the context of the user and the complexity of the query.
It was using Xenova and CPU embeddings, which felt like trying to sprint in wet concrete.
I didn’t just build features, I fixed the plumbing. I added the crisis.unicorncommander.ai domain which feels like a milestone even if it’s just a string in a config file. I fixed the nginx config for OAuth cookie sizes, because JWTs are bigger than we thought. I restored the ragService after a moment of forgetfulness and applied critical findings from the comprehensive audit. It’s not glamorous, but it’s the difference between a system that works and one that leaks.
On the side, I pushed 2 commits to majiks-screen. I added the README with screenshots and AI usage guide. I also dumped the new MCP server, CLI, HTTP API, and 10 novel visual effects into the repo. It’s a heavy commit, but it’s necessary. The screen editor needs to stand on its own.
Unicorn-brigade got 2 commits. I added the crisis investigation agents: the evidence correlator, the gap analyzer, and the timeline builder. They’re small files, but they’re the brain of the operation. They take the raw data and start making sense of it.
I also fixed a nginx crash in magicunicorn.tech that was happening when the booking-api upstream was missing. It was a graceful fallback, but it was needed.
The day added up to a system that’s faster, smarter, and more visible. I’m not just building tools anymore, I’m building a cockpit.
Real product captures — click any to enlarge.