EvidenceGap AI

EvidenceGap AI Making document reviews clearer, faster, and easier to trust

*Inspiration

I work with project documents regularly, and one thing I've learned is that finding a file is often only half the problem.

Contracts, invoices, supporting records, and other documents may be scattered across different folders. Even when everything seems to be there, checking whether the information matches can take a surprising amount of time.

I've worked on organizing large collections of project documents, dealing with inconsistent filenames, deeply nested folders, and records that weren't always easy to identify. That experience made me think: what if a NAS could do more than just store these files? What if it could help us understand which documents support each other, where information conflicts, and what still needs attention?

What it does

EvidenceGap AI is a read-only document review Skill designed for UGREEN AI NAS.

A user selects a folder or specific documents and describes what they want to check. The Skill then uses the available NAS capabilities to locate relevant files, extract information, and compare the evidence.

Its goal is to answer practical questions:

• Do these documents contain conflicting information?
• Can the findings be traced back to their sources?
• Which supporting documents were not found in the selected folder?
• Are there files that could not be reliably examined?
• What still needs someone to review it manually?

Rather than presenting every AI-generated answer as a fact, EvidenceGap distinguishes between confirmed findings and uncertain results.

It also leaves the original files untouched.

How I built it

I developed the initial Skill using OpenAI Cod alongside UGREEN's official UGOS Skill Creation Tool.

The Skill is built around UGREEN's documented capabilities:

• ug-filesystem.find for discovering files within an authorized folder.
• ug-document.metadata and ug-document.extract for examining documents and retrieving relevant information.
• ug-document.qa for comparisons supported by source citations.
• ug-document.reconcile for supported contract and invoice or receipt amount reconciliation.

I also designed an Evidence Coverage Matrix to distinguish verified documents, possible matches, evidence not found within the searched scope, unreadable files, and results that could not be verified.

An Exception Register makes it easier to understand which findings still need human attention.

The entire workflow is designed to remain within the folders or files explicitly authorized by the user.

Challenges I faced

One of the biggest challenges was designing something useful without assuming the platform could do more than its documented capabilities allow.

For example, finding a file with a particular name doesn't necessarily prove that it's the correct document. Similarly, failing to find something in one folder doesn't mean it doesn't exist somewhere else.

I had to account for incomplete searches, ambiguous filenames, documents belonging to different projects, and situations where the available evidence simply isn't enough to reach a conclusion.

Another challenge is testing. At this stage, I don't have access to a UGREEN AI NAS runtime, so I focused on creating a Skill that follows the official development standard rather than claiming functionality I haven't tested.

What I learned

This project reinforced something I had already experienced while working with real document collections: good automation isn't just about getting answers quickly. It's about knowing where those answers came from and recognizing when they might be incomplete.

I also learned more about designing agent workflows, working within a documented capability system, and turning an everyday administrative problem into a reusable AI Skill.

The official UGREEN static checker passes the current Skill without reported errors or warnings. Its behavior has also been reviewed against synthetic document scenarios, although those scenarios have not yet been executed on NAS hardware.

What's next

My next goal is to test EvidenceGap on an actual UGREEN AI NAS environment, verify how well document extraction and citations work, and refine the workflow based on real results.

If the Skill performs as intended, I would love to explore ways to make evidence reviews even more accessible to small teams that handle large numbers of documents.

For me, EvidenceGap AI is a first step toward making everyday document work less repetitive, more transparent, and easier to manage.

Built With

• ai-agents
• openai-codex
• ugos
• ugreen-nas
• uliya