This week, the noise floor dropped. For a long time, I’ve been building systems that talk, but rarely systems that understand. I was chasing a specific kind of coherence. I wanted the numbers to match, the billing to be honest, and the interface to stop feeling like a placeholder. By Sunday night, the arc was clear. We had a tax engine that actually files, an accounting ledger that traces its own lineage, and a brand system that finally looks like a product. The audit came back clean, the moat got deeper, and for the first time in months, I didn't feel like I was holding water with a sieve.
It started on Monday with a fresh surface. I dropped the initial commit for Brand-Ops, 22,449 lines across 136 files. It was the whole stack, from the Dockerfile to the MCP integration. It felt good to have a new product surface in the world. But the real movement happened in Meeting-Ops. I shipped v3.29.0 through v3.30.0 in a single push. This was the audit remediation wave. I tightened Brigade tenancy to fail-closed, added WebSocket handshake auth, and closed the audit’s last security item. I tore out speaker avatars that didn't belong and replaced Whisper references with the actual Parakeet stack. The Sessions speed got an uplift, and I closed out summary idempotency so the finalize path doesn't re-invent work. Contact-Ops got a knowledge graph that populates by default, powered by a force-directed 2D layout. The overview clusters by employer, and the dashboard counts are now real. Ops-Center got a security overhaul I’m glad to have behind me. I implemented a default-deny auth middleware, which meant going back through every endpoint and asking who actually needed access. I locked down 11 residual admin endpoints and stopped writing user PII to the request log.
Tuesday was a tangle. I chased three threads at once, and they all knotted together. Contact-Ops was the big one. I shipped the relationship graph end to end, A to Z directory, with 3D visualization by default. The graph was stubborn, though. Three was throwing tick-of-undefined until I deduped three and d3-force-3d. That was six commits just to make the 3D scene stop being blank. Behind the graph, I baked in a Data Quality cockpit and a review-queue triage cockpit. I wired the sovereign local Qwen 3.6 35B as the analyst in the chat pipeline. Ops-Center got a billing overhaul with per-org gateway-key metering. I built a never-lose-money pricing safeguard and closed a live unauthenticated data leak with a default-deny auth middleware hotfix. Project-Ops went deep on tenancy. Phase 2 of workspace isolation shipped across the board. I scoped PATs to workspaces at the app layer and wired auth login to return a workspaces array. The iOS app was the heaviest commit count but also the noisiest. I built on-device summarization with a chunk-map-reduce pipeline and wired the MLX Gemma runtime. The iOS app was shaping up to be a real v1 candidate.
I dropped the initial commit for Brand-Ops, 22,449 lines across 136 files.
Wednesday started with something that refused to boot. I fixed tax-planning-ops so a fresh deploy actually boots on real Postgres 17. The Dockerfile was copying a nonexistent requirements file. The config aliases were wrong. Five fixes, one deploy that worked. After that, I had enough momentum to keep building. The tax engine got its real teeth today. Phase 1 finally wired up LiteLLM with qwen3.6-35b-a3b for 1040 extraction. The UI got a complete redesign from scratch. Money displays in dollars, not cents. The tax math and OBBBA constants are correct, and the append-only audit chain with metadata-covered hashes is in place. Center Deep Research got spun up from nothing. Initial commit on the sovereign deep-research engine and canvas. Then the space dashboard, plan review, agent CRUD, live choreography, and security hardening landed in rapid succession. Meeting-Ops shipped v3.36 with portable voiceprints and same-meeting duplicate detection. The transcriber was broken by a single enum change. Fixing it reminded me how one character can sink a week.
Thursday was a heavy day. Ninety-one commits across ten repos. The arc was clear: I was building the machinery to bill federated inference correctly while shipping the tax-planning state engine. The tax-planning-ops work was the biggest chunk. The state income tax engine is config-driven now, with 19 states live including the big brackets in CA, NY, NJ, and VA. The comprehensive return orchestrator folded SE tax and the §199A QBI deduction into the main calculation path. I spun up a full C-corp engine that handles 1120, QSBS, and AET. Multi-year projections are wired through. The ops-center federation work was a different kind of grind. I spent a lot of time making sure the billing side of federated inference didn't double-debit credits. There was a metering reconciliation pass where every call now emits exactly one Lago event, no matter how many hops it takes through the federation layer. Trust modes became runtime-enforceable. The unknown-peer default was made flippable via platform settings. The accounting-ops web side got a Living Ledger dashboard v2 behind the NEXT_PUBLIC_DASHBOARD_V2 flag. Real data instead of stubs.
Friday was about connecting the pipes and polishing the glass. The headline is the brand-ops release. GUI v2 is live. I spent the day pushing the final polish and fixing the surface-level bugs that QA round 1 unearthed. The suite SSO is now fronted by oauth2-proxy, and I fixed the proxy cookie lifespan to match the realm’s SSO session. It is the brand as interface now. Behind the scenes, the accounting-ops repo took a heavy hit. I activated the Qwen Tier-2 categorization via the metered gateway. I also built a frontier-fallback routing policy so the system escalates to a better model if the first pass abstains. The tax-planning-ops repo got a lot of attention for the federation with accounting-ops. I built the trusted M2M caller path so TPO can read P&L data from Accounting-Ops. This involved passing the workspace_id through the Brigade exchange and brokering tokens via RFC-8693. I also made sure the net income matches the contract exactly using Decimal-exact dollar strings. The strategist engine now has deeper context, pulling in client goals, prior-year trajectory, and the full return memo.
Saturday was split between two worlds. The first was listing-ops, where the vision pipeline hit its wall on RTX 6000 context. I squeezed image sizes, capped OCR tokens, and bumped per-slot batch from four to two. It was dependency hell in the best way. Every fix exposed the next, but by the last commit dots.ocr was grounded, the text tasks were on their own route, and the whole thing actually fits now. The second world was bigger and more interesting: tax-planning-ops v3. I had a design spec that mapped every surface from the Desk down to Season, then spent the day building it. Daylight and Paper token system, dark mode that actually works, a primitive SVG chart kit, routing scaffold. Then the swarm ported in waves: Book, Brief, Playbook, Case File, Explained, Staff, Return, Plan, Entities
Real product captures — click any to enlarge.