Blanco Brand 三人核心團隊如何在 25 個項目實現五分鐘部署
Five-minute deploys across 25 projects with a core team of three
Blanco Brand 將部署時間由一小時縮短至五分鐘,平台承載 25 個項目,由三人核心開發團隊維護,服務可用率為 99.95%。團隊採用 DigitalOcean App Platform 等託管服務、重複使用相近的後端架構,並以 CI 測試把關變更。AI 模組透過 OpenRouter 接入語言模型,另設值班、警報及監控應對快速部署帶來的風險。
An hour-long deploy doesn't cost you an hour. It costs you every deploy you decide not to do.
At Blanco Brand a deploy used to take an hour. Now it takes five minutes. The platform holds 25 projects on DigitalOcean (deployments of different sizes, not 25 products) for clinics, mining companies, residential communities and a few other sectors, it runs at 99.95% uptime, and the core development team is three people. I'm the CTO and I still write code there.
People ask what tool makes that possible. I don't have one to name. We made most of the stack boring on purpose, so our attention goes to the parts that can't be. Below are the five rules we follow, and what each one costs.
Why an hour is worse than it sounds
A lot of my work happens next to the people using the software, the job that now goes by Forward Deployed Engineer. It goes like this: someone at the clinic shows you a screen, says "this field is in the wrong place", and you can fix it in ten minutes. If shipping the fix takes an hour, it goes into a batch for "later". The batch grows, a big batch is scary to ship, and nobody volunteers to ship scary on a Friday.
At five minutes, the fix ships while the person is still at their desk. SuiteMedica, our clinic SaaS, logged 1,468 commits from three authors between 31 March and 16 November 2026. That rhythm would be painful with hour-long deploys.
Rule one: don't run what you can rent
With three people at the core, nobody on the team should be the person who babysits a database host.
Our services deploy on DigitalOcean App Platform, which also runs 9 managed databases for us (PostgreSQL, MySQL and Valkey). Errors go to Sentry. Identity on many of the newer systems is AWS Cognito, and files go to S3.
Nothing exotic. Every piece we rent is one that can't page us for an OS patch, a disk filling up or a backup script that silently stopped.
Separately, development spend at Blanco Brand fell 63% after we reorganised delivery around a stable team, and infrastructure spend fell 55%. I'm not going to credit those numbers to any one tool. But I'd make the same infrastructure call again tomorrow.
The price is control over the box. For software that runs clinic schedules and lab certificates, I'll trade it for sleep.
Rule two: one backend shape, repeated
Look across the projects and the same skeleton keeps showing up:
- Alex Stewart (mining inspection lab): NestJS API on MongoDB, Cognito, S3, React with Vite.
- HemoSuite (haemodynamics back office): NestJS on MongoDB, Cognito, S3 and SES, React with Vite.
- CCDM Omni and CCDM Derivaciones (two systems for Clínica Ciudad del Mar): both NestJS, MongoDB and Cognito.
- MED360 Presupuestos: NestJS on MongoDB, Cognito, BullMQ on Redis, Socket.IO, React with Vite.
- CARS Control: NestJS 11 with MongoDB as a read cache, Cognito, React 19 with Vite 7.
When the shape repeats, a lot of things stop being decisions. Auth works the way it worked last time. In the NestJS projects, queues are BullMQ and real time is usually Socket.IO. Someone who has worked on one of these repos can open another and know where the modules are. On a small core team that matters more than any framework benchmark, because any of us should be able to fix any of them.
Boring doesn't mean uniform at any cost, though. SuiteMedica runs NestJS on PostgreSQL with TypeORM and a Next.js web app. Keeper's API has been in production since 2019 on Node.js with Fastify and Express, and its admin panel is Angular. I took over as Keeper's CTO in February 2024 and left that stack alone. A rewrite to make the repos match would cost real work, and no customer would notice the result.
Rule three: the pipeline decides, not a person
Five-minute deploys are only safe if "deploy" isn't where the judgement happens. The judgement goes into CI.
On SuiteMedica, GitHub Actions runs, before any change is accepted:
- Build
- Type-check
- Lint
- Unit tests
- End-to-end authorization tests
The last one matters most. In SuiteMedica every clinic is a tenant, and authorization is the most delicate part of the system. A bug that shows one clinic's data to another clinic is worse than any outage. So it gets its own end-to-end suite, separate from the unit tests, and it runs on every change.
The repo has 100 test files. HemoSuite has 137, and CARS Control has 149 on Vitest. These counts aren't a quality score. They show that in these projects tests get written alongside the code, not promised for later.
What I'd tell another small team: put your scariest failure mode in CI first. For us that's cross-tenant access. For you it might be money, permissions or a migration. Coverage numbers can wait.
Rule four: keep the volatile parts behind a seam
AI is the part of the stack that changes fastest, so it's the part I keep most isolated.
SuiteMedica's booking agent answers patients over WhatsApp, phone, Instagram, Messenger, email and web chat, and puts the appointment in the clinic's calendar. It lives in its own package in the monorepo, agente-citas, and it reaches language models through OpenRouter, so the model can change without touching the rest of the code. CARS Control also reaches models through OpenRouter for its AI module.
Running the agent itself is its own story (AI agent in production). For deploys, the lesson is short: if something changes every month, give it a boundary you can swap behind.
Rule five: fast deploys need someone awake
Fast deploys move risk. They don't remove it.
Ship in five minutes and, sooner or later, you'll break something in five minutes. The patient messages don't stop at 11pm because a deploy went out at 10:55. So SuiteMedica has an on-call rota, alerting and monitoring, and Keeper and CARS Control report errors to Sentry. The upside is symmetric: the fix also ships in five minutes.
Before you make deploys faster, decide whose phone rings at 11pm. If I had to rank the five rules, I'd put this one first. A pipeline without an owner at night is just a faster way to find out in the morning.
Fewer decisions, on purpose
Nothing here is advanced engineering. Each rule takes a decision off the table, so a small core team can keep 25 projects in their heads.
Whoever sets the standards has to live with them in the same repos, which is why I still write code as a CTO (more on that here). The SuiteMedica case shows these rules in one real system.
If you run a small team across a lot of projects, I'm curious what you put in CI first. Tell me in the comments, or write to hola@bcristi.com.
來源:DEV Community · dev.to



