Local document intelligence / human review

OficIA

A local document-intelligence workbench that turns sensitive PDFs into source-backed proposals for human approval.

At a glance

What I owned

Period: July 2026
Role: Product shell, workflow integration, and browser acceptance
Team: Joaquín Fernández Cedeira on the core pipeline; Franco Ferreira on OficIA UX v2 and the real-data product interface
Users: People reviewing sensitive operational documents before an accountable decision is recorded
Status: Live auth-gated system; local Document AI with universal human review

Contribution

Designed and implemented the OficIA UX v2 product shell, connected the existing audited queue and review workflows, clarified field-level evidence states, and recorded desktop and mobile acceptance.

Tradeoff

Let local inference remove repetitive extraction work while keeping the source, corrections, and final operational decision visibly owned by a person.

Evidence

Authored implementation commits and QA records in the canonical repository, a live health endpoint and auth boundary, and an earlier public LLM Local PDF baseline. All portfolio documents and values are synthetic.

Deployment: Local-first inference
Workflow: Every document reviewed
Evidence: Field-level anchors
Record: Versioned human decision

Documents become review-ready work

OficIA turns an incoming PDF into a prioritized review item, not an autonomous action. The inbox separates operational documents from controlled tests and makes ownership, source quality, and missing evidence visible before a reviewer opens the file.

Synthetic OficIA inbox prioritizing documents for review while separating operational work from controlled tests — recreatedA review inbox organizes operational documents without mixing them with pipeline tests.

Review without leaving the source

The workbench keeps extracted instructions, the original document, and the decision panel in one context. Reviewers can navigate the source, correct proposed values, and choose the next action without rebuilding the document in another tool.

Synthetic OficIA workbench showing instructions, original PDF context, proposed values, and human decision controls — recreatedInstructions, source, and decision stay together throughout review.

Every value earns its place

Each critical field carries a source snippet, page location, and visible validation state. Missing or conflicting support stays attached to the field, so reviewers can open the exact fragment that justifies or challenges a proposal.

Synthetic OficIA field-evidence chain connecting a proposed value to its source snippet, page anchor, and validation state — recreatedField-level evidence makes every proposed value reviewable against its source.

A human decision becomes the record

Approval is not the end of the interface. Corrections, versions, reviewer identity, evidence state, and the final action become an auditable record before anything is published to the next system.

Synthetic OficIA version history showing corrections, reviewer identity, evidence state, approval, and controlled publication — recreatedThe final record preserves what changed, who decided, and which evidence supported publication.

Project linksGitHub