From the README

# ETL-D MCP Server 🌊

A deterministic data middleware for AI agents, exposed as a standard Model Context Protocol (MCP) server.

## The Problem: The "Data Tax" in LLMs

Asking an LLM to parse a 10,000-row CSV, a Spanish Norma 43 bank statement, or an EDI 850 file directly in the context window leads to three major issues:
1. **Token Exhaustion:** Sending raw B2B formats to a context window is highly inefficient.
2. **Hallucinations:** LLMs struggle with strict spatial alignment and counting. A single misplaced comma in a financial trade history is catastrophic.
3. **Non-Determinism:** You cannot build reliable B2B pipelines if the extraction format changes based on the LLM's "mood."

## The Solution: Waterfall Architecture

ETL-D stops agents from attempting to "read" raw B2B data. Instead, the agent routes the file/string to this MCP server, which processes the data using a strict **3-Layer Waterfall Architecture**:

1. **L1 - Heuristic (0-Shot):** Standard Python parsers (`regex`, `