Reqfold

Inspiration

The lowest supplier quote is easy to spot. The lowest quote that actually matches the request is harder.

We started with a narrow question: what happens when a buyer compares price before reconciling the technical commitments scattered across an RFQ and supplier responses?

Reqfold makes that reconciliation visible. It is not another procurement chatbot. It is a sourcing-review workbench built around a requirement, its source, the supplier's quoted value and a repeatable check.

What it does

Our synthetic demonstration follows an order for 500 CNC-machined brackets and three supplier quotes.

The €10,900 quote looks cheapest. Checking six requirements reveals two conflicts—tolerance and delivery—and one missing certificate commitment. The €11,650 quote is the lowest that matches all six listed requirements.

The distinction matters: missing evidence is a clarification request, not proof that a supplier cannot provide it. Likewise, a quoted certificate commitment is not authentication of an issued inspection certificate.

Every matrix result opens an evidence view. A buyer can inspect the requested value, quoted value, source excerpt and comparison rule, then draft a clarification or export the complete review record. Nothing is sent or purchased automatically.

How we built it

The application uses dependency-free HTML, CSS and JavaScript, deployed on Vercel. A server-side endpoint calls a configured open-weight model through Featherless to extract a typed requirement ledger. The existing integration was exercised successfully against the production endpoint during development.

Numerical bounds, calendar dates and explicit certificate-list membership are checked by deterministic JavaScript—not another LLM judgment. The interface exposes live extraction separately from an explicitly labeled sample-ledger mode. An unavailable model call does not silently become a successful AI run.

The final release adds a side-by-side supplier matrix, source inspection, a quote-revision simulator, clarification drafts and JSON review export. Model text is escaped before display, and unsupported or malformed evidence remains unresolved.

Challenges and what we learned

Our hardest decision was where to stop trusting generated interpretation and start requiring explicit evidence.

A second challenge was proving the result was not hard-coded. In the demo, changing the cheapest supplier's tolerance, delivery date and certificate commitment makes the same verifier select that quote. The simulated amendment is visibly marked as a what-if, not a real supplier response.

We added 16 release checks for baseline results, revisions, missing evidence, invalid numbers and dates, unsupported rules, optional conditions and certificate matching. Local Chromium tests covered the interactive workflow, evidence dialogs, revision and restore, export, error display and mobile layout.

Why this use case

The prototype is designed for technical buyers working across supplier documents rather than inside a replacement procurement platform. Its European manufacturing setting is demonstrated through the sourcing scenario; we do not claim that open-weight hosting alone guarantees European data residency.

Prototype scope and next steps

The current version uses fixed synthetic RFQ excerpts and pre-normalized quotes. It does not parse arbitrary PDFs or CAD files, connect to ERP systems, authenticate certificates or award contracts. Prices assume the same quantity, currency and commercial basis.

The video is an explicitly labeled sample-ledger walkthrough of the running interface. The deployed Extract requirements button is the separate live Featherless path.

Next steps are extraction evaluation against real consented documents, robust ingestion, source-change tracking, access controls and deployment-specific data policies.

AI interprets. Code checks. People decide.

Built with: JavaScript, HTML5, CSS, Vercel, Featherless.ai Live demo: https://reqfold.vercel.app/ Repository: https://github.com/bitrocks/reqfold