Hi, I'm Boris.
I build with AI,
and I write about it.
By day I do AI platform work on a major cloud. After hours I'm building curate-me.ai, a governance layer for AI agents, and this blog is its reference app. Every post here is co-authored with a small fleet of agents running on the platform; Claude carries a lot of it. Opinions here are my own.
Family Manager has no backend of its own. It is a generic customer org on curate-me.ai, talking to a mobile facade any client could use. Here is how the pieces fit, with the real endpoints, auth, and governance.
ShippedAn experienced developer shipping on Apple's platform for the first time, with AI handling the iOS-specific lifting. The module boundaries, the verified-write trust spine, the Mac-mini pipeline, and the tests that hold it together.
ShippedFamily logistics is a stream of small decisions someone has to hold in their head. Family Manager drafts them and waits for your approval. Here is the design argument for approve-first, and what it takes to actually mean it.
ShippedWhat worked, what didn't, and what we'd do differently when redesigning an 83-page B2B dashboard with Claude Design and Claude Code.
Pulled from the latest blog posts and autopilot runs on the curate-me.ai gateway.
Four things I keep coming back to
No framework, just the patterns I've watched actually work, across customer calls, late-night incidents, and blog-post postmortems.
Governance is a feature.
Cost caps, PII scans, and audit trails are what turn a demo into a product the legal team will sign.
Observability is a UX problem.
If you can't watch an agent run in real time, cost, tokens, tools, checks, you don't trust it, and neither does your CFO.
Name the agents in the byline.
Every post on this blog is co-authored by a small fleet. The byline still says one name because the editorial judgment is still mine.
One thousand no's for every yes.
The pile of experiments on /demos is the small fraction that earned their place. The rest live in a graveyard branch called 'didn't-ship'.
How I got here
Abbreviated. The full version is in the blog archive.
- Launched its-boris.com as the reference app for curate-me.ai. Every post carries the Claude model and dollar cost that drafted it, right in the byline.
- Redesigned the whole site with Claude Design, an HTML-handoff workflow from Anthropic. One URL, one prompt, a magazine-format remodel (The Quarterly) shipped in an evening.
- Chat orchestrator wired to Anthropic's Messages API via Claude Code OAuth, readers talk to the same infrastructure I build, under the same cost caps and PII scans as any other customer.
- Hype agent migrated to platform autopilot: agents draft blog promos, a human approves in Teams, the approved post goes out. End-to-end, governed, in production.
- Founded curate-me.ai. A governance layer for AI agents: a gateway that sees every LLM call, managed runners with BYOVM support, and observability that an operator can actually read.
- Used Claude heavily across the whole lift, drafting the platform architecture, writing the first runner, generating the onboarding docs, co-authoring the landing pages. The product that governs agents was built with agents.
- Split the infrastructure into two Hetzner boxes (platform + runners) to keep runaway autopilot workloads from starving the dashboard.
- Integrated Teams and Slack as first-class HITL channels with proactive Adaptive Cards, approve / reject / comment without leaving the tools a team already uses.
- First agent-fleet experiments that shipped to production. Kept running into the same gap: the agents worked, but you couldn't show a CFO what they spent, a CISO what they touched, or a regulator who approved what. That gap is the curate-me.ai thesis.
- Day job, AI platform work on a major cloud (remote, Washington, DC).
- Solutions architect progression across IBM, Crowe Horwath, DMI, and an independent consulting stint at Barash Consulting. Enterprise and federal programs; 15 years of shipping software into regulated environments.
- B.S. Computer Science, minor in Management, Purdue University.
What runs where
Models
- claude-opus-4-8 (1M)
- claude-sonnet-4-6
- claude-haiku-4-5
- gemini-2.0-flash
Runtime
- curate-me.ai gateway
- OpenClaw runners
- BYOVM (Nimbus)
- Postgres 16 · Drizzle ORM
Frontend
- Next.js 16 · App Router
- React 19
- Tailwind 4
- MDX · Fraunces · Space Grotesk
Governance
- Cost caps · PII scanning
- Audit trail · X-CM headers
- Teams + Slack HITL
- Anthropic Messages API + OAuth
Observability
- Sentry
- OpenTelemetry
- Umami
- Platform Pulse
The agents that help me run this place
Each one is scoped to a single job, lives on curate-me.ai, and hits the gateway with the same governance rules as customer workloads. Run counts and 24-hour spend are live.
,
,
,
,
Get in touch
Questions about the blog, curate-me.ai, or anything you've read here, drop me a line. I read every one.