From 40 solicitations a week by hand to an eight-step automated bid pipeline

A five-person certified small-business reseller supplying office, janitorial and facilities products to state and local government agencies ran its entire bidding process manually — reviewing 20 to 40 solicitations per week by hand.

Every step was human. Pull solicitations from a subscription portal. Read each one to decide whether it was product-only, since any labour or installation component disqualifies the business completely. Score it. Create a folder, download attachments, read PDFs and spreadsheets to pull out line items. Price each item against at least three vendors. Build a pricing sheet, fill an internal purchase-order workbook, complete the agency's compliance forms, assemble a bid package, and track it all across a legacy tracker and a set of spreadsheets. Roughly half the target solicitations don't even carry their documents in the source portal, forcing a manual hop to the agency's own site.

We built a daily pipeline that takes a solicitation from just-posted to ready-for-a-human-to-sign, in eight steps.

It pulls newly posted product solicitations each morning and decides whether each is worth bidding — checking it is a pure product purchase with no labour, that it falls in the client's categories, that the deadline is far enough out and the value in range — then scores it 0–100 and files a card either way, including the rejects, each with a written reason.

For those that pass it creates a dated folder, downloads the agency's documents, and reads them to extract the actual shopping list — item numbers, descriptions, part numbers, quantities, units — whether that list sits in a PDF table, a spreadsheet, or buried in prose. Structured parsing runs first and the model is the fallback, so most extraction costs nothing.

It then prices every line item against live distributor data feeds, and where no part number is specified proposes the three best-fit products with a similarity rating each. Prices flow into two spreadsheets cloned from the client's own templates, with the profit margin left as an editable cell rather than baked into the formulas. Finally it generates a cover letter and company information page, auto-fills whatever agency forms are machine-fillable, flags those needing notarisation, merges everything into a single bid package PDF, and opens a tracker task with 32 populated fields that doubles as the bid cover sheet.

Submission stays deliberately manual — a person reviews and files.