GreenLedger

GreenLedger turns loan agreements into living ESG systems — extracting, tracking, and verifying sustainability obligations with AI. Built for the LMA Edge Hackathon 2026, it's a privacy-first desktop application where all processing happens locally.

The problem in financial services

Banks issue loans with ESG clauses — requirements around carbon reporting, environmental targets, social impact metrics, governance standards. These obligations are buried in dense legal documents. When compliance deadlines pass unnoticed, institutions face regulatory risk, reputational damage, and financial penalties.

How GreenLedger works

Upload a PDF loan agreement and GreenLedger does the rest. The document goes through a multi-stage pipeline: text extraction, chunking, and AI-powered ESG clause extraction using a keyword-triggered approach — keywords flag potential clauses, then the local LLM analyses and categorises them.

Extracted clauses appear in a structured interface where compliance teams can review, approve, or reject each one. Approved clauses feed into the compliance dashboard, which tracks obligations, deadlines, and status across the entire loan portfolio.

The RAG chat interface

Beyond structured extraction, GreenLedger includes a natural language chat interface powered by RAG. Ask questions like "What ESG targets does this loan require?" and get answers with mandatory source citations — every claim references a specific document and page.

Anti-hallucination guarantees

For financial compliance, accuracy isn't optional. GreenLedger uses five layers of protection: rule-based keyword triggering before LLM involvement, mandatory citations on every RAG answer, explicit refusal when context is insufficient, confidence scoring that flags low-confidence extractions for human review, and deterministic compliance calculations with no LLM in the math.

Privacy-first architecture

The entire system runs locally. The LLM (Llama 3.2) runs via Ollama on the user's machine. ChromaDB stores embeddings locally. SQLite holds structured data on disk. No document content ever leaves the user's computer — critical for handling confidential loan agreements.

GreenLedger was built in a hackathon timeframe, but the architecture is designed for production. The privacy-first approach and anti-hallucination guarantees make it suitable for real financial services use cases.