PythonAgentic AIGeminiGoogle APIs

Tenderer Eligibility Document Parser

An agentic-AI document pipeline that reads eligibility documents with Gemini, validates its own output through five quality layers, and processes 100+ documents a day — converted from an n8n workflow into production Python.

Overview

Tender eligibility documents are messy and high-volume, and pulling the right sections out of them reliably is exactly the kind of judgement task that breaks rigid rule-based parsers.

This microservice treats Gemini not as a plain API call but as an autonomous agent: it discovers files, reasons about each document, validates its own extraction, and only then commits to the database — tracking every step in real time. It is a backend-only microservice of EZBid, a procurement-intelligence platform.

What I Built

I converted an existing n8n visual workflow into production Python, orchestrating four Google services (Drive, Sheets, Gemini) plus MySQL through a state-machine pipeline with a dedicated validation engine.

Agentic AI extraction — Gemini performs context-aware analysis and reasons about output quality before anything is persisted.

State-machine pipeline — documents flow NEW → IN_PROGRESS → COMPLETED/ERROR with real-time Google Sheets tracking.

Five-layer validation with hallucination detection and truncation prevention before any database commit.

Multi-service orchestration — Google Drive discovery, Sheets tracking, Gemini processing, and MySQL persistence behind clean service adapters.

Resilient recovery — exponential backoff (1s→16s) and up to five retries, achieving a 99%+ success rate at 30–60s per document.