Inspiration
Deals get agreed out loud: "GST is included", "we'll take the old unit away". Then the quote says "+GST" and nothing about removal. The voice memo and the quote often both sit on the same NAS, but nobody lines them up until the bill arrives.
What it does
Point it at a recording (a voice memo, voicemail or walkthrough video) and a document (a quote, lease or contract). It transcribes the recording on the NAS, translates it if the recording and the document are in different languages, extracts the written terms and labels every spoken promise: matches, contradicts (with the dollar difference), or no written term found. Every flagged row is read a second time against the exact transcript segment and document page before it is reported. The result is a mismatch sheet with timestamps and both wordings quoted, contradictions first, saved next to the document.
How we built it
A Markdown-only orchestration Skill on the Uliya standard. It makes nine delegated steps across file-agent, knowledge-agent and document-agent: find, transcribe, save the transcript, translate (only when languages differ), extract, compare, double-check flagged rows, save the sheet, and optionally tag. Each step uses only the real upstream result. A row check drops any comparison row whose timestamp or quoted term is not in the real transcript or extraction, and a second read confirms every contradiction against the exact segment and page. Everything runs locally, and writes and translation go through platform approval.
Challenges
Keeping the comparison honest: a model can invent a mismatch. So every row must point back to a real transcript timestamp and a real extracted term, flagged rows are read twice, and the sheet says "not found in the document" rather than implying anyone acted in bad faith.
Accomplishments
The Skill passes a local static check that mirrors every rule in the kit's checker, with 0 errors and 0 warnings. It also handles every delegation status, empty and partial results, a consent gate and repeat use.
What we learned
The third status, "no written term found", is the most useful one. No tool that reads a single file can produce it.
What's next
Real NAS testing in Round 2: transcription and translation accuracy on accents and noisy job sites, how often the double-check overturns a row, and the revised-quote rerun.
Built With
- markdown
- ug-document
- ug-filemgr
- ug-filesystem
- ug-knowledge
- ugos
- ugreen-ai-nas
- uliya-skill-kit
