Inspiration
A lot of people keep contracts, invoices, and years of project files on their UGREEN NAS — freelancers, small studios, home-lab owners. The files are irreplaceable, but they're also scared to let AI touch them. One bad rename or a misread document and the folder structure is wrecked. And sending client documents to a cloud AI? That's a privacy non-starter.
We kept thinking: what if you could try the automation on a copy first? Like a canary in a coal mine, but for your files. That's where GhostTwin came from — bring staging, verification, and rollback to NAS workflows.
What it does GhostTwin makes a verified clone of your NAS folders, lets AI loose on the clone, and only touches the real files if everything checks out — or rolls back in one click if it doesn't.
Fingerprint first. Before anything changes, we inventory the folder and hash every file with SHA-256. That's our baseline — we know exactly what "untouched" looks like. Build the twin. We copy the files into an isolated staging folder and re-hash them to prove the copies are identical. Live files never get modified. Let the AI audit the twin. Photos of receipts get OCR'd, PDFs get reconciled against each other, and everything lands in a plain-English verification summary — all inside the sandbox. Save the evidence. Every hash, duplicate, and diff gets written into an audit dossier you can actually open and read. You decide. Nothing promotes to live without your approval. If the audit looks good, results go live; if anything's off, one click restores everything to the exact baseline. How we built it Followed UGREEN's Uliya skill standard (v2.0.0) with the official skill kit, targeting the full main-agent profile on UGOS. The interesting part was orchestration: the main skill can't call tools directly, so we built a chain of four specialist agents — file, photo, document, knowledge — each handing verified outputs (file IDs, real paths, hashes) to the next. Nothing downstream ever sees a guessed path. Split the spec into four reference docs so the main skill stays readable. Passed UGREEN's official validator with zero errors and zero warnings on the first serious pass. Challenges we ran into The security model fights you. The parent skill is banned from touching the filesystem itself — everything must go through delegate calls with single-quoted instructions. Getting hashes to flow correctly between agents took more iterations than the AI parts did. Real folders are messy. Empty directories, corrupted PDFs, permission-denied files — we had to make every step fail safe: stop the chain, report exactly how much was covered, never widen the scope on its own. Accomplishments we're proud of Zero-error validation. check-skill.mjs came back clean — every structural and security rule passed. It actually works across modalities. A photo of a receipt, a PDF invoice, and a folder of duplicates all reconcile into one audit — that's something only a local NAS with unified storage can do privately. The rollback is real. Every restore is verified against the SHA-256 baseline, so "undo" isn't a hope — it's a checksum. What we learned Local AI on a NAS is a different beast than a chatbot. You can't just fire commands — you need hard boundaries: look before you touch, get approval before you change, and always be able to prove you can undo it. Once you accept that, the cross-modal stuff (receipt photo ↔ invoice PDF ↔ audit trail) becomes genuinely useful.
What's next Real hardware test on the iDX6011 Pro — staging, OCR, reconciliation, and rollback on actual multi-GB folders during Finalist judging. More evidence types — meeting recordings via audio transcription, delivery photos via surveillance events, folded into the same audit. Publish to the Skills Hub with presets for freelancer deal rooms, tax-season receipts, and media cleanup.
