Inspiration
Visa, scholarship and job interviews are short, and the officer has your papers in front of them. Under pressure it is easy to say "forty thousand" when the bank letter says 34,500, or the 20th when one form says the 25th. Many people practise by recording themselves, but few replay a recording with every paper open next to it. Those papers are passports and bank statements, which shouldn't be uploaded to a chatbot.
What it does
You record a practice interview and keep your papers in one folder on the NAS. You ask Uliya: "Check my visa interview practice against my Papers folder." The Skill:
- finds the recording and the papers;
- transcribes the recording and asks which voice is yours;
- lists the facts you stated (dates, amounts, schools, programmes, employers, sponsors, travel) for you to confirm;
- looks each one up in your papers, including photographed letters, comparing like with like (same subject, same period, same currency);
- marks each fact Matches, Differs, Papers disagree, Not in your papers or Unclear, with the file and page;
- if a fact is in none of your papers and you ask, searches your photo library's text index for a photo of the missing paper (it only lists leads);
- saves the review as a note, only after you say yes, so the next rehearsal shows what you fixed.
It never suggests what to say. If a paper is wrong, it tells you to get the paper corrected. It is for your own preparation, not for judging other people.
How we built it
It is an orchestration-only Skill: SKILL.md plus seven Markdown references, built with the official uliya-skill-kit 0.7.3. Every NAS action is delegated to a registered expert:
file-agent(ug-filesystem, file-workflow) finds the recording, the papers and earlier reviews;knowledge-agent(ug-knowledge) transcribes and saves the note;document-agent(ug-document) answers each fact from the PDFs, Word files and photos of papers, with evidence;photo-agent(ug-photo), on request, searches the photo text index for photos of missing papers.
Every delegation handles completed, failed, cancelled, waiting_user and waiting_approval. Steps 1 to 9 only read; the one write is a new note after the user's yes, never overwriting. A repeated request is recognised before transcription, so it costs nothing and creates no duplicate. ID numbers are shortened to their last four characters. The Skill calls no outside service; UGREEN says transcription and document reading run on the NAS by default, and with the NAS's local AI model the recording and papers never leave the box. The package passes the kit's check with zero errors and zero warnings.
Challenges
- Deciding what counts as a "fact" and when two values really match: Fall 2027 vs 20 August 2027, "about thirty-one thousand" vs 31,200, a yearly amount vs a one-off total, 03/04/2027 on a US form vs a Pakistani one, "fifteen lakh", and names whose spelling speech cannot show. These rules are written out in
references/claim-rules.md. - Papers that disagree with each other. They got their own result, because then no spoken answer can match both.
- Doing the duplicate check before the expensive steps rather than after.
What we learned
The kit's checker is stricter than its docs. For example, the root may only hold SKILL.md and references/, and capability lines need a specific arrow format.
What's next
Testing on a real NAS: speaker labels in transcription, and page numbers from the document expert. Urdu recordings once NAS transcription supports Urdu (today it accepts English, Chinese, Japanese and German). Then a mode that compares two partners' recordings for family visa interviews.
Built With
- markdown
- ug-document
- ug-filesystem
- ug-knowledge
- ug-photo
- ugos
- uliya
- uliya-skill-kit