I design systems end to end, then run them in production.
baryo.dev · @BaryoDev · writing · email

14 years in production software. I design systems end to end and then run them: an event-sourced multi-tenant CMS, the deploy tooling it ships through, the libraries underneath it, and the front end on top. Strongest in .NET, though most of what I ship is not.

The decisions are the work. Event sourcing over CRUD and what that costs on the read side. Modules compiled in rather than loaded at runtime, because a plugins folder is a place where writing a file runs code. One cheap VM instead of a platform, and what that trades away. Each of those is written up with the reasoning, not just the outcome.

@BaryoDev is 30 public repositories across C#, TypeScript, Go and JavaScript, all open source and self-hostable. No paid tier, no seat cap, no metered anything.

Eleven merged fixes and five accepted bug reports in libraries I run in production, across four repos totalling 13k stars. Each was found by using the thing, not by browsing for issues. The badges are live searches, so they answer for themselves.

Most of these reports were fixed by the projects' own maintainers rather than by me, which is the part worth pointing at: the report carried enough evidence for someone else to act on it.

Sometimes the useful contribution is not a pull request. On marten#5457, someone else's leak report, I could confirm half of it and disprove the other half, and said so rather than opening a fix I could not demonstrate. The maintainer's own #5466 opens by crediting that comment: "right on both counts and shaped this."

upstream-fixes writes each one up, and carries a runnable reproduction for marten#5215 against published NuGet packages, so you can watch that bug appear on one version and disappear on the next rather than take my word for it.

Open-source headless CMS for .NET. Event-sourced on Marten and PostgreSQL, multi-tenant, user-defined content schemas, 13 opt-in modules, an event-triggered workflow engine, and a Next.js admin UI.

127 endpoints and 2,284 tests, run against real PostgreSQL and MinIO through Testcontainers. On .NET 10 and Marten 9. MPL-2.0, and every module is included rather than sold separately.

4.0.0 shipped in September 2026 to NuGet, GHCR and Docker Hub: four auth and file-access holes closed, GDPR erasure, tenancy resolved from a domain, content references, multi-architecture non-root images, an SBOM, and a WCAG pass.

The claims are gates rather than sentences. CI restores a backup, upgrades a 3.x database, applies the Kubernetes manifests, generates the SBOM, scaffolds a module from the template and builds it, resolves every compose file, and refuses an image tag that is amd64-only.

Live at playground.baryo.dev/barakocms. Console: barakoBrew, for designing content types, roles, workflows and integrations against the API. Typed client: barako-client. baryo.dev itself is rendered from it.

lean-agent-method. How I put AI coding agents through a real ticket backlog without burning a month of plan in one night, plus the script that measures cost per change.

A cheaper model drafts every change. Scripts, not instructions, run the mechanical checks. A cheap critic reviews every change against six fixed questions, and the expensive model only sees what the critic cannot close. At most four changes in flight.

That took me from roughly 66 dollars of model use per change to about 25, with no drop in what got caught. Those are my numbers, at list prices, on one .NET codebase, so treat them as a shape rather than a benchmark. The method and the measuring script are both in the repo.

30 public repositories under @BaryoDev. Each solves one problem, ships to a package registry, and is tested rather than described.

Mostly postmortems of my own mistakes, and decisions with the reasoning attached.

More at baryodev.medium.com.

I sort decisions by how expensive they are to reverse. A storage engine, a tenancy model, a public contract and a licence are one-way doors: those get argued out in writing first, with the reasoning kept rather than just the outcome, so the next person can disagree with the argument instead of guessing at it. Most other things are cheaper to try than to debate. That is why barakoCMS 4.0 shipped one-way doors only and everything else waited for 5.0.

Tests that cannot fail are the defect I look for first: a mock returning what the real dependency never returns, an assertion satisfied by a type's default value, a benchmark that prints and exits zero. Before trusting a gate I break the thing it guards and check it goes red.

AI-assisted daily, held to the same gates as everything else. The failure mode I design against is a green suite that proves nothing, which is why the mechanical checks are scripts and every change gets a critic before it gets an expensive model.

Languages C#, TypeScript, JavaScript, Go, SQL Architecture Event sourcing and projections, multi-tenancy (conjoined and database-per-tenant), modular monoliths over plugin runtimes, API contracts and versioning, licensing and one-way-door calls Backend .NET 8 to 10, ASP.NET Core, REST and GraphQL, microservices, CQRS-influenced design Messaging RabbitMQ, SQS, SNS, webhooks, background jobs Data PostgreSQL, SQL Server, Marten, Redis Cloud AWS (CDK, ECS, Fargate), Azure (Functions, App Services, DevOps), Oracle Cloud, Docker Testing xUnit, NUnit, Moq, Testcontainers, Playwright, Vitest, BenchmarkDotNet, security scanning in CI

📫 arnelirobles@gmail.com · 🌐 baryo.dev · ✍️ baryodev.medium.com