Arvelis / AI application engineering
AI that ends
in working
software.
I built an application engine around structured definitions, deterministic validation and explicit human control.
Laravel / PHP · React / TypeScript
Inside the contract / metadata, not generated code
A proposal becomes
something inspectable.
Entities, typed fields, relations and workflows form a definition that the runtime can understand.
Deterministic validation
Structure is checked.
Meaning still needs judgment.
The validator checks supported types, references and constraints. An unsupported field cannot pass as an accepted definition.
A supported structure can proceed to human review.
The decisive checkpoint
Human review.
Before creation.
Validation cannot decide whether a definition represents the real operation. A person owns that decision.
Explicit confirmation is part of the product workflow.
Accepted metadata → application runtime
One engine.
A working application.
Forms, records and controlled actions come from the accepted definition. The neutral product interface below is real; no client workspace is invented.
What the implementation demonstrates
One engine.
An explicit contract.
- Model, not generated code
- Entities, typed fields, relations and workflows are metadata. The runtime supplies forms, records and controlled actions.
- Change is reviewed too
- Workspace evolution uses a proposed patch, validation and human confirmation before transactional application.
- Boundaries stay visible
- Structural validation does not prove business correctness. Semantic review remains human work; arbitrary customer code is not executed.
An engineering experiment, not evidence of customer adoption. The neutral application screens shown here demonstrate the product entry and review workflow; no populated client workspace or business outcome is claimed.
What the work proves