TenderCopilot

A discovery feed, automatic extraction of requirements from PDFs via Google Document AI, a drafting workflow with Gemini, a content library, and billing.

OutcomeReads the tender PDFs the founder used to comb through by hand.

Brief

Greek companies lose hours searching for tenders on ΚΗΜΔΗΣ and ΕΣΗΔΗΣ and reading 80-page PDFs to pull out the technical requirements. They needed a tool that understands Greek government documents and separates what matters from the procedural noise.

Work

• Backend in Next.js 14 (App Router), Prisma + Postgres, auth via Auth.js (Google and email)
• A discovery feed with filters (CPV, value, deadline), saved searches, and flagging for false positives
• A PDF analysis pipeline: Google Document AI → structured requirements → Gemini for classification
• A worker queue for asynchronous parsing and monitoring
• A content library of reusable blocks, tagged and searchable
• A tender page with tabs: overview, requirements, draft, tasks, analytics
• A billing surface via Stripe with usage tiers
• An e2e suite via Playwright (smoke, auth, and regression, across 3 browsers)

Highlight

Tuning Document AI for Greek was the real puzzle. The out-of-the-box models missed structures typical of Greek tenders. Custom post-processing and a Gemini second pass reached 90%+ parse confidence, with flagging for human review where needed.