← Back to projectsEspañol
Case study · Original work

From requirements to a reviewable technical design

I designed and developed an AI-assisted engineering workflow that connects requirements, code evidence and design proposals with automated checks and human review.

José Antonio Gallegos Carvallo · Architecture, development and automation

The problem and my contribution

Understanding existing behavior and preserving validations require more than plausible generated text. I designed the workflow, developed its tools and simplified a multi-agent process to one agent while retaining deterministic checks and human approval.

Workflow and architecture

  1. 1. Requirements

    Prepare the case.

  2. 2. Evidence

    Read code through MCP in read-only mode.

  3. 3. Design

    Link proposals to source fragments.

  4. 4. Validation

    Check document structure and references.

  5. 5. Human review

    Review and approve the proposal.

  6. 6. Word

    Generate and verify the approved document.

The agent, MCP integration and read-only knowledge base feed evidence and a Markdown proposal into local validators, human review and document generation.

Architecture decisions

Fictional example

A reservation with a valid date

A fictional application must reject confirmation if the date is missing or before the minimum allowed date, while preserving its existing availability check.

Fictional existing code (original identifiers retained):

101: Event 'Confirmar'
102:    ValidarDisponibilidad()
103:    GuardarReserva()
104: Endevent

Check the reference to GuardarReserva():

Finding: line 108 is not present in the evidence.

This browser demonstration checks one exact citation; it does not compile or functionally test the proposed implementation.

The public demonstration uses invented data and code; the original system is not exposed.

Evaluation examples

Supported response

Check that the response matches its source.

Unsupported claim

Flag a statement absent from the evidence.

Out of scope

Check that missing information is acknowledged.

These are proposed evaluation scenarios, not measured model results.

Results and limitations

A pilot completed the document workflow with human review and Word generation. Document checks do not prove functional behavior. Time and token savings require comparable measurements; compilation and functional testing belong to implementation.

Open fictional sample document
← Explore more projects