ESG COMPLIANCE AGENT: AI-POWERED REGULATORY ASSESSMENT

Overview

A production ESG compliance assessment system that uses AI to analyze company sustainability documents against ESRS (European Sustainability Reporting Standards). The system features a local LLM for completely offline operation, making it suitable for confidential enterprise environments.

Processes a full ESRS assessment across all 10 modules in under 3 minutes.

Architecture

┌─────────────────────────────────────────────────────────┐
│           26 DATA SOURCES                                │
│   News APIs · Regulatory feeds · Company reports        │
│   Market data · Carbon monitoring · ESG databases       │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│           COLLECTORS (Selenium + API clients)            │
│   Automated scraping · Rate limiting · Deduplication    │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│           FAISS VECTOR INDEX (BGE embeddings)            │
│   Semantic document search · Context retrieval          │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│           RAG PIPELINE → LLM (Granite 1.3B, 4-bit)     │
│   Per-module analysis · Evidence extraction             │
│   Compliance scoring with uncertainty propagation       │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│           COMPLIANCE REPORT (HTML)                       │
│   10 ESRS modules × score (0-100) + evidence + gaps    │
└─────────────────────────────────────────────────────────┘

The system follows a 4-layer architecture with SOLID principles:

• Presentation — Interactive CLI with Rich terminal UI and 5 context modes
• Application — Command registry, process orchestration, report generation
• Domain — Core models, interfaces, ESRS compliance logic
• Infrastructure — RAG server (FastAPI on port 8001), data collectors, API clients

ESRS Coverage

Each module receives a compliance score (0-100) with uncertainty propagation:

• E1 Climate Change — Scope 1/2/3 emissions, transition plans
• E2-E5 — Pollution, Water, Biodiversity, Circular Economy
• S1-S4 — Workforce, Supply Chain, Communities, Consumers
• G1 — Business Conduct, Ethics, Governance

Example Assessment Result

Module: E1 Climate Change
Score:  78 / 100  (±6 uncertainty)
Time:   ~15s

Findings:
  ✓ Scope 1 & 2 emissions disclosed (high confidence)
  ✓ Transition plan referenced in sustainability report
  ⚠ Scope 3 partially covered — supply chain gaps
  ✗ Science-based targets not yet validated by SBTi

Key Features

• Local LLM — Granite Tiny 4.0 (1.3B params, 4-bit GGUF) for zero-network inference
• RAG Pipeline — FAISS vector store with BGE embeddings for semantic document search
• 26 Data Sources — News APIs, regulatory feeds, company reports, market data, carbon monitoring
• 10 ESRS Modules — Full coverage: E1-E5 (Environmental), S1-S4 (Social), G1 (Governance)
• Automated Reporting — Professional HTML compliance reports with per-module scoring

Tech Stack

Python, FastAPI, LangChain, FAISS, sentence-transformers, llama-cpp-python, Selenium, Rich, pytest