Javier Casal

Javier Casal is a software engineer whose work spans web applications, shared frontend packages and AI-assisted engineering workflows. At Flywire, that work includes Orbit, an internal platform he described jointly with Eduardo Simon: ordinary code handles predictable operations, while models perform bounded tasks inside reviewed workflows. A code-review workflow makes the distinction concrete—the model can read code but cannot modify it, and human review remains part of the process.

Web applications and readable code

Javier’s earlier work included a joint project with David Cogollos Ribera at the Universitat Politècnica de València: migrating a commercial sports-management application from a client/server architecture to the web. The university included their collaboration among its recognized industry-linked projects for 2006–2007.

In his June 2011 exploration of Clean Code, Javier examined Robert C. Martin’s guidance through concrete naming decisions. Names should reveal intent, distinguish concepts meaningfully and spare readers unnecessary mental translation. His focus was the work required to understand and maintain a program after it has been written.

His public code includes Spring MVC and Spring Data JPA examples, an Angular demonstration and an experiment with F# type providers. His npm publishing activity also includes flystyles, Flywire’s UI framework, and @flywire/react-flywire-js, its React integration package. That activity documents participation in shared company tooling, rather than sole authorship of those packages.

Models inside reviewed workflows

In their joint account of Orbit, Javier and Eduardo describe workflow definitions written in YAML and reviewed through the same merge-request process as other code. Each model step declares the tools and access it can use; structured outputs determine subsequent branches. This puts decisions about when to invoke a model, what it may access and what happens afterward into an inspectable workflow.

The code-review example illustrates how those restrictions support a recurring engineering task. The speakers reported more than 9,700 review runs in 30 days using a model with read-only access to code. Other applications include translation updates, dependency fixes, ticket triage and bulk repository changes. Orbit separates orchestration, execution and session state, with timeouts, rate limits and locks to prevent duplicate runs.

The speakers also distinguish operational execution from answer quality. New workflows initially receive a few checks against real cases, golden datasets run on demand, and online evaluations remain under development. Full sandboxing is a goal rather than an established capability. Their reported execution success therefore does not establish that a model’s answers are correct. The adoption approach they describe starts with a useful skill, proves a recurring task and then moves it into a background workflow with explicit ownership and review.