-
-
Proofline turns a cited evidence run into a fingerprinted decision packet while preserving the human approval boundary.
Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for Proofline
Inspiration
AI agents are increasingly good at gathering facts and drafting documents, but business risk concentrates at the final step: the moment a recommendation becomes an approval, signature, or payment. Proofline makes that boundary visible. It lets automation do the reversible work while preserving an accountable human decision.
What it does
- A reviewer enters an organization and the decision under consideration.
- SerpApi runs focused searches across official-company, adverse-event, and current-news queries.
- Proofline classifies the returned evidence, preserves source URLs, and calculates a run fingerprint.
- Doctavian receives the structured run data and merges it into a decision-memo template.
- Nutrient DWS turns the auditable HTML packet into a deterministic PDF.
- The interface presents the recommendation, evidence ledger, fingerprint, and explicit human approval boundary.
How we built it
Proofline is a React and TypeScript application deployed on an edge-compatible runtime. The evidence route calls SerpApi and normalizes organic results into a compact evidence ledger. A document route prepares the same decision payload for two sponsor paths: Doctavian uploads JSON, merges it into a DOCX template, and streams the generated document back; Nutrient DWS receives an audited HTML packet and performs the PDF build.
A SHA-256 digest binds the organization, decision, evidence, and generation timestamp to the displayed approval state. The application also provides a conspicuously labeled deterministic demo state so it remains testable without ever presenting simulated evidence as live.
Sponsor API proof
SerpApi
SerpApi performs the current-intelligence work. Three targeted Google searches retrieve registry context, adverse-event language, and operating news. Proofline preserves the source links and classifies the actual result text. The production route is live. A verified production run returned seven current results from three targeted searches while preserving source links and classifying the actual result text.
Doctavian
Doctavian performs the structured document-generation path. Proofline uploads the completed run as JSON, merges it into a purpose-built decision template, requests PDF output, and streams the returned artifact. The integration route and DOCX template are in the public repository; sponsor credentials have been requested and can be activated through environment variables without changing the flow.
Nutrient DWS
Nutrient DWS is live in production. It performs the deterministic review operation by converting Proofline's exact audited HTML packet into a PDF. We verified the public endpoint end to end with a fresh run and a valid PDF response.
Challenges
The hardest design problem was not producing another risk score. It was preserving provenance across multiple API boundaries while making the stop-here-for-a-person moment obvious. A second challenge was designing a credential-safe demo: missing sponsor keys must never silently turn simulated results into claims of live retrieval.
Accomplishments that we are proud of
- One coherent product uses three sponsor APIs for substantive work rather than decorative calls.
- Evidence sources, risk signals, timestamps, and the packet fingerprint remain visible to the reviewer.
- The application is publicly accessible and the Nutrient PDF path is verified in production.
- The decision packet includes a real human sign-off area instead of implying that an agent approved the transaction.
- The public repository includes the complete application, environment template, and Doctavian merge template.
What we learned
Multi-API agent workflows become more trustworthy when every stage has a narrow contract: search returns evidence, generation returns a document, deterministic processing returns a review artifact, and only a person returns approval.
What's next
Next steps are configurable policy packs by industry, saved evidence snapshots, reviewer authentication, and change detection that invalidates approval when any cited source or generated clause changes.
Try Proofline
Built With
- api
- cloudflare
- doctavian
- javascript
- node.js
- nutrient
- react
- serpapi
- sha-256
- typescript
Log in or sign up for Devpost to join the conversation.