The groundwork
A machine reads the paperwork. People decide what it means.
A tender is rarely one document. It is a cover sheet that links to the files where the real requirements live, and those files are where bids are lost. Reading all of them by hand takes an afternoon per tender, so most bidders do not. Our pipeline does it in minutes — and the division of labour is deliberate: fixed rules handle every number, because a model that misreads ₹48 lakh as ₹48 crore will do it fluently.
- 01
Fetch
Tender documents are pulled from the portal itself, along with every file they link to — scope of work, buyer-added conditions, price formats.
- 02
Read
PDF, Word and spreadsheet are parsed to text. Scanned pages, which have no text layer at all, are read by a vision model rather than skipped.
- 03
Extract
Every figure that can disqualify a bidder — turnover, EMD, passing marks, deadlines — is pulled by fixed rules, never generated. Each carries the text it came from.
- 04
Judge
Only then does a model reason over the facts: what conflicts, what is buried in an annexure, what decides the award. Every claim quotes its source.
Every extracted figure carries the page and the sentence it was read from, and the report says plainly what it could not read. That is the whole standard: checkable, not confident.