Tuesday, 58 commits across 12 repos. Glitter Mane got the bulk of it, and honestly, that's the headline.
I spent the day building out the personal medical companion stack inside glitter-mane, which has been growing into something that feels genuinely useful rather than toy-like. The command center took shape across a dozen commits, but it coalesced around four real threads.
First was the imaging pipeline. I wired up TorchXRayVision as the first real imaging model, running as a CPU compose service, then built an ensemble architecture on top of it for zero-shot triage and change detection. The whole thing uses a model registry that can swap in 2026 SOTA serving recipes without breaking anything. It's still local, still on-prem, which is the point.
I spent the day building out the personal medical companion stack inside glitter-mane, which has been growing into something that feels genuinely useful rather than toy-like.
Then came the diagnostic process itself. I added history-taking with self-critique, guideline grounding that makes consults cite real clinical references instead of making stuff up, and a deterministic medication-safety net that checks interactions, cautions, and duplicates against local JSON files. The holistic whole-record review ties it all together, doing the cross-domain look a rushed visit can't give.
But the thing that took the most commits was security work. Three separate review swarms hit the codebase, and I patched everything from two critical findings up through low-severity stuff. The adversarial review alone caught two critical and five high. The second swarm found a high, three medium, and five low. There's a pattern here, and it's the same one I always see with medical AI: the fun stuff is the models, the real work is making sure nothing breaks in a way that hurts someone.
Also did voice consults with on-prem Parakeet STT and on-device TTS, family/dependents management under one account, FHIR R4 export to complete the import/export loop, and a daily check-in tracker. The whole thing is becoming a real product.
On the email-ops side, I shipped the agent registry and autonomy dial, which is the fleet identity layer. Each agent gets a stable key, an autonomy level, and a sending mailbox. The agent-inbox approval queue UI wired up for human-in-the-loop review of outbound drafts. Also added the email-steward health surface with REST, a UI panel, and MCP tools. And I fixed a workspace-scope bug in the pending counts that came up during adversarial review.
The Microsoft 365 connector in customer-ops landed as dormant by default, which is the right call. It's a real module with OAuth, token cipher, health seam, and a dashboard settings card, but nobody's forced into it until it's proven. I also swapped the customer-ops brand mark to the constellation icon across all the logo files.
Brand-ops got Iris a face. The director character finally has a visual identity, plus an off-by-default hosted-generation entitlement gate. The M365 docs got written up too.
The smaller repos kept their rhythms. unicorn-brigade got live event streaming mid-loop via asyncio queue, real token metering for streamed chats, and a Three.js performance win from reusing node objects. The SSE ticket moved into an HttpOnly cookie instead of the URL, which is the right security move. unicorn-stable's RAG got hybrid retrieval with dense and sparse indexing plus contextual chunking that grounds each chunk in its document before embedding.
union-ops got a full Pocket Steward rewrite so it's actually usable end-to-end, plus a public landing page with request-access intake. listing-ops capped AI titles at 80 characters and made vision batches run concurrently. majiks.music required a BYO ACE Music key for generation. majiks.online added EU AI Act Art. 50 AI disclosure labels on tracks. ops-center renamed The Colonel from Col. Corelli because the period was annoying everyone. accounting-ops got a tax-document OCR pipeline that classifies and transcribes through a local vision VLM.
Fifty-eight commits and most of them were in one direction: making the medical AI thing real, secure, and actually usable. The email fleet is getting its bones right too.
Real product captures — click any to enlarge.