Shipyard

PromiseCheck

Paste a borrower's reply from a collections call to see if it is a real promise to pay, how likely it is kept, and what to do next.

Did you make PromiseCheck?

Claim this listing to edit it or take it down. It's free.

Claim this listing
Screenshot of PromiseCheck

Reactions from the wild

1 reaction.

  • Shipyard member24 Oct
    Interesting scope: classifying a pasted reply into promise strength and a suggested next action. Setup friction worth checking is how it handles ambiguous or multi-sentence replies with mixed signals (partial payment offer plus excuse) and whether it needs call metadata or transcript formatting, or just raw plain text. Does it return a confidence score alongside the likelihood estimate, or just a category label?

About PromiseCheck

A prototype for one question: when a borrower tells a collections voicebot "haan, kar dunga", is that a real promise to pay?

Most voicebots log every yes as a promise to pay. PromiseCheck reads what was actually said, pulls out the promise (date, full or split, hedges, someone else paying, whether the bot put the date in their mouth), scores how likely it is to be kept, and picks the next action: confirm by message, firm it up with a callback, or hand it to a human. Every planned call stays inside the RBI 8 am to 7 pm window.

The live demo takes a typed or spoken reply and can run a 30-case evaluation. The call data is simulated, and the README says so: the rules extractor scores 19 of 30 on a hand-written golden set of hard cases. Built with Lovable and n8n, with an LLM extractor prompt for Gemini.

Summarised by Shipyard from the app's public site and repository.