Inspiration
Teaching an AI works a lot like teaching a student. It studies a pile of flashcards (pictures with names on them, like "fox" or "owl"), then sits an exam on pictures it has never seen. The exam score tells you how good it really is.
But things go wrong quietly:
- Exam answers end up in the study pile. If some exam pictures were also used for studying, the AI has already seen the answers. Its score looks great, but it isn't real.
- Cards have the wrong name. A fox photo labelled "owl" teaches the AI the wrong thing.
- Old attempts pile up. Every study session saves a large copy of the AI. Nobody remembers which one was best, and the disks fill up.
This happens to experts too. In CIFAR-100, a famous picture collection used to test AI, 1 in 10 exam pictures were already in the study pile. Once they were removed, scores dropped by 9–14%.
People who build AI often keep all of this work on a NAS. A UGREEN AI NAS already holds the files and has its own AI built in. So we asked: why not check the whole project right where it lives?
What it does
Sift is a quality check for the AI you build, right on your NAS. You ask once, and it runs three checks:
- Data check: finds exam pictures hidden in the study pile, the same picture filed under two names, pictures that may have the wrong name, and names with far fewer pictures than the rest.
- Training check: reads the notes each attempt left behind, shows which attempt really worked best, and warns about attempts that crashed, can't be repeated, or have a score that isn't real.
- Storage check: keeps the best and latest saved copy from each attempt, and shows how much space you can get back by archiving the rest.
The checks help each other. Problems in the data show which scores can't be trusted, and the best attempt decides which copies are worth keeping.
Nothing changes until you approve it. Sift can tag pictures for review, move old files into an archive folder and save the report. It never deletes anything. Run it again after more training and it shows what changed.
How we built it
Sift is a Skill for Uliya, the assistant built into UGREEN AI NAS, made with UGREEN's official Skill Developer Package. Sift is a written plan, not an app. It hands each job to one of the assistants already inside the NAS:
- File assistant: maps the project, finds copied files, and archives files after approval
- Photo assistant: uses the NAS's own AI to look at a small sample of pictures for each name
- Document assistant: reads the notes each training attempt left behind
- Knowledge assistant: saves the report next to the project
Sift only looks until you approve a change, and it passes along only real results from one step to the next. No passwords, no cloud, and nothing leaves the NAS.
Challenges we ran into
- Explaining AI problems in plain words. The hardest part was making the report make sense to anyone, not just AI experts. That's where the study-pile and exam wording came from.
- Only promising what the NAS can do. The NAS finds exact copies, not resized ones, so Sift says that clearly in every report.
- Keeping it fast. Checking every picture with AI would be slow, so Sift looks at a small sample for each name and marks anything it flags as "please review", never "wrong".
- Never guessing. Every number in the report comes from a real result. If one check fails, it is marked and the others still run.
Accomplishments we're proud of
- The full Skill passes UGREEN's own checker with zero errors and zero warnings.
- Three checks that work together, in one conversation.
- A design that is useful without ever deleting anyone's files.
What we learned
How the NAS splits work between its built-in assistants, how to write a Skill that hands off each job cleanly, and how much clearer a product gets when you explain it to someone who has never trained an AI.
What's next
Sift hasn't run on a real UGREEN AI NAS yet. If we become finalists, we'll test it on a practice project with 3,000 pictures and planted mistakes: 40 exam pictures hidden in the study pile, 15 wrong names and a crashed attempt. We'll check that Sift finds every one, then tune it with real results. After that, we'll add projects beyond pictures.
Built With
- markdown
- ugos
- ugreen-ai-nas
- uliya
- uliya-skill-kit
