ETL-D – AI 에이전트(MCP 서버)에 대한 결정적 데이터 구문 분석

ETL-D is a deterministic data middleware designed to act as a shield for AI Agents. It stops LLMs from "hallucinating" over structured data by providing a strict, 3-layer parsing architecture via the Model Context Protocol (MCP). Standard LLMs are terrible at reading raw B2B files (CSV, PDF, EDI, Norma 43). They suffer from: - Token Exhaustion: Sending a 5,000-row CSV to context is a waste of money. - Precision Loss: LLMs struggle with spatial alignment. A misplaced comma in a bank statement is a financial catastrophe. - Non-Determinism: You can't build a reliable pipeline if the output format depends on the LLM's "mood." The ETL-D engine processes every request through a strict hierarchy: - Layer 1 (Heuristic): 100% Python-native deterministic parsers. 0% Hallucination risk. ~70ms latency. - Layer 2 (Semantic): Column-to-Schema alignment using embedding-based routing. - Layer 3 (LLM Shield): Strict JSON enforcement using Llama 3.3 70B with additionalProperties: false constraints for high-entropy noise. - Open your claude_desktop_config.json . - Add the ETL-D server (Ensure you use version 3.3.3 ): { "mcpServers": { "etld": { "command": "npx", "args": [ "-y", "@pablixnieto2/[email protected]" ], "env": { "ETLD_API_KEY": "YOUR_API_KEY_HERE" } } } } - Option A: Manual Purchase (Human) Get a Starter Pack ($5 / 5,000 credits) or a Subscription at api.etl-d.net. - Option B: Zero-Touch Provisioning (Agent-Led) 🤖 If you don't have a key, simply ask Claude: "I don't have an ETL-D key, can you help me get one?". The agent will call the /provision tool, generate a Stripe Checkout link for you, and automatically set up the key once paid. Zero-touch, human-in-the-loop. - Financial & B2B Heavy Lifting parse_bank_statement: Support for Spanish Norma 43 (N43). Turns raw bank files into clean JSON. parse_trade_history: Deterministic extraction of trades, fees, and dividends from complex broker exports. parse_edi: ANSI X12 EDI parser (Optimized for 850 Purchase Orders). generate_sepa_xml: JSON to PAIN.008 (Direct Debit) XML generator. - Document Intelligence pdf_to_spatial_markdown: Crucial for Agents. Converts PDFs to Markdown preserving table structures before the LLM reads them. extract_invoice / extract_resume: High-accuracy schema extraction for standard B2B documents. - Atomic Enrichment (1 Credit/call) enrich_amount: Cleans "Total: 1.240,50€" into {amount: 1240.50, currency: "EUR"}. enrich_date: Resolves human-readable dates ("next Friday at 5pm") with Timezone awareness. enrich_address: Standardizes global messy addresses into structured components. accounting_map: Maps concepts to ES PGC, US GAAP, or IFRS. Cloud Engine: Hosted at api.etl-d.net (Python/FastAPI). Python SDK: pip install etld. n8n Nodes: Available in the n8n community as n8n-nodes-etld. MIT - Created by Pablixnieto2