Inspiration
Whenever something goes wrong at home, like a scratched car, a parcel that "was delivered" but never arrived, or a landlord holding back a deposit, the proof usually already exists. It is just scattered: a camera event in one app, phone photos backed up somewhere else, a scanned policy in a folder nobody has opened in a year. Under stress, people miss items, send undated photos, and weak claims get rejected.
At work I build multi-agent systems and RAG solutions, so I kept seeing the same pattern: the hard part is not finding one file, it is putting the pieces together and being honest about what is missing. A UGREEN AI NAS already holds all of those pieces, privately, in one place. That made it the right home for this idea.
What it does
ProofPack is a Skill for UGREEN AI NAS (Uliya). You describe the incident and roughly when it happened, for example "Build a claim pack for my car scratch last Tuesday night." ProofPack then:
- Confirms the claim type and time window before searching anything.
- Searches camera events, photos and documents for that window, read-only.
- Extracts the facts an insurer, landlord or seller will ask for, translates foreign-language documents, and shows what changed between versions (for example a revised repair quote).
- Builds one timeline where every line names its source file.
- Checks each required item and marks it Verified, Single source, Conflict, Weak or Missing.
- Flags photos that show other people, so bystanders and children can be left out.
- Only after you confirm, creates a case folder, an evidence album and a cited summary note.
It never moves, deletes or shares your files, and it never decides who is right. If the courier says "delivered 14:32" but the door camera shows nothing, you see both facts side by side.
The core rule is simple. For a checklist item $i$ with sources $S_i$:
$$ \text{status}(i) = \begin{cases} \text{Missing} & |S_i| = 0 \ \text{Conflict} & \exists\, a, b \in S_i \text{ that disagree} \ \text{Verified} & \geq 2 \text{ independent sources agree} \ \text{Single source} & \text{otherwise} \end{cases} $$
Weak is applied on top, using rules for each claim type, such as damage photos taken long after the incident.
How I built it
This was a solo build. I started by writing out several possible concepts and scoring each one against the four judging criteria: user value, innovation, technical quality and AI NAS fit. ProofPack scored highest. I then folded in the best ideas from the others: conflict detection, version comparison and the bystander check.
I built the Skill with Claude Code and the official UGOS Skill Creation Tool (uliya-skill-kit):
- SKILL.md orchestrates five expert agents (Surveillance, Albums, Files, Documents, Knowledge) over ten steps. Each step passes real event IDs, photo IDs and file paths to the next, never vague references.
- references/ holds five claim recipes (vehicle damage, delivery, warranty, rental deposit, home damage), the evidence rules, output formats, a 25-case failure table, a test plan and disclosures.
- I ran "Help me check this skill" repeatedly until no required changes remained.
The package is Markdown only: no scripts, no credentials and no external downloads. All test data is synthetic.
Challenges I ran into
- No NAS in Round 1. I had to design against the documented capability scope without running anything, so I wrote every step to depend only on listed capabilities and planned the device tests in advance.
- Not overclaiming privacy. My first draft said everything ran locally. Then I realised Uliya itself may use a cloud AI service, depending on the user's settings. The Skill now sends nothing off the NAS itself, and asks before reading documents so the user can decide.
- Designing for absence. The easy failure for an AI tool is to fill a gap with something similar. I had to make "nothing found" a first-class result: name what was searched, mark it Missing, and stop rather than substitute.
- Fitting it into 500 words. Cutting the proposal down forced me to decide which features really mattered.
Accomplishments that I'm proud of
- The five-status evidence check. It tells you not just what was found but whether your case actually holds together.
- Conflicts are shown as two facts, never as a verdict.
- Nothing is written until you say yes, and even then only new folders, copies, an album and a note. Originals are never touched.
- The Skill passes the UGOS Skill Creation Tool check.
What I learned
- In evidence work, honesty is the feature. A tool that says "this is missing" is more useful than one that confidently fills the gap.
- Confirmation gates and failure handling take as much design as the main flow. The failure table grew to 25 cases.
- Privacy claims need to be exact. "Runs locally" and "this Skill sends nothing itself" are different promises.
What's next for ProofPack
- Round 2 testing on a real UGREEN AI NAS: confirm the surveillance output format, person search, album creation and reading accuracy on scanned documents.
- A demo video that walks through a full vehicle-damage pack and a conflicting delivery claim.
- More claim recipes, such as medical bills and travel disruption, once the core flow is proven.
- Better handling of poor scans, so unreadable fields are flagged clearly for a manual check.
