I work at the points where simple systems stop being simple.
Case studies, projects, and technical writing on scale, failure, legacy systems, and business rules. Each piece shows the constraint, the decision, and where the solution stops.
Method
I start with what happens when a message fails.
The database cannot change. The legacy core cannot stop.
The happy path comes later.
The decision records the discarded alternative and the condition that would justify revisiting it.
Code shows what was built; the case shows why.
Proof
~98% fewer intermittent failures in critical flows
Intermittent failures between microservices with no contract for retry, DLQ, or duplication. Adding more consumers only pushed the overload elsewhere.
Publicação confirmada, consumo com prefetch controlado, retry com backoff e DLQ por fluxo. Escalar só consumidores fica de fora do desenho.
~7 min → <1 s commissions, from the original load to the warm cache
The database belonged to another team. The dashboard had to stop depending on a schema we did not control.
Cada estágio corresponde a uma decisão incremental documentada no case. Números de outras histórias não entram neste desenho.
~25% fewer tables in the domain-based split
The product could not stop, and the original authors were already gone. Before migrating, we had to discover which boundaries the database still revealed.
O banco legado revela fronteiras; pessoas, autenticação e aeronaves saem gradualmente para contextos com persistência própria.
Artifacts
Public repository
Importing CSV must not block the Nest API. The campaign persists and publishes to RabbitMQ; consumers in TypeScript, Go, or Rust record the result in Mongo. The API does not wait for the carrier.
CSV na API Nest com persistência Postgres; publicação na fila; consumidores no mesmo contrato gravam no Mongo. A API não espera a operadora.
Documented experiment
Maestro (Codex/Cursor) delegates over MCP; FIFO daemon and OpenCode workers execute with state in SQLite. Orchestration worked, but measurement did not confirm cost or time reduction.
Maestro delega; daemon FIFO coordena um worker OpenCode por vez; SQLite guarda estado. A hipótese de eficiência não se confirmou.
Own product, desktop
Profile and immutable versions stay in the machine's SQLite. The supervised agent only proposes changes under schema; ATS and PDF/DOCX export reuse the same graph, with no SaaS backend owning the data.
Perfil e versões no SQLite; candidatura no mesmo contexto; agente só propõe sob schema; ATS e export reutilizam o grafo sem SaaS dono.
Path
2025 to 2026
Rental and fleet platforms. The public proof from this period is the messaging redesign.
Infosistemas
2023 to 2025
SQL Server owned by another team. The dashboard had to work without relying on a schema we did not control.
VBET
2023
PHP 5.7 still ran the operation. Automation had to free support without rewriting the monolith.
Maxmilhas
2022 to 2023
White-label so banks could join without a per-client fork.
South System / QUIQ-Itaú
2021 to 2022
The monolith had no original authors. Migration started with the parts of the database we could still read.
Flapper
Practice
Each step constrains the next. Without those constraints, architecture becomes personal taste.
01
Who is hurt when this breaks, and in which operation.
02
What cannot stop, change, or be treated as extra capacity.
03
The path taken, with the discarded alternative kept visible.
04
What was measured, in which scenario, with whose authorship.
05
The condition that would justify revisiting the decision.
