Back to Projects
Sustainability Tech · Seed Stage

ESG Portfolio Risk Analysis with Critique-Correct Agents

The client needed to audit thousands of unstructured ESG reports to identify high-risk investments, a process previously requiring weeks of manual analyst labor.

Critique-Correct Loop LangGraph Claude 3.5 Sonnet Pinecone
Business Impact
92% reduction in report processing time

The Problem

A sustainability-focused investment firm was drowning in ESG documentation. Their analysts spent weeks manually reviewing thousands of unstructured reports—PDF sustainability disclosures, annual reports, third-party audits—trying to identify “greenwashing” and quantify actual environmental impact. The manual process was not only slow but prone to inconsistency.

The Architecture

flowchart TB
  subgraph intake [Document Intake]
      PDFs[ESG Reports / PDFs]
      Unstructured[unstructured.io]
  end
  
  subgraph agents [Critique-Correct System]
      Extractor[Extractor Agent]
      Auditor[Auditor Agent]
      Reconciler[Reconciliation Engine]
  end
  
  subgraph memory [Knowledge Layer]
      Pinecone[(Pinecone Vector DB)]
      SourceDocs[Source Document Store]
  end
  
  subgraph output [Output]
      RiskReport[Risk Assessment Report]
      Flags[Greenwashing Flags]
  end
  
  PDFs --> Unstructured
  Unstructured --> Extractor
  Extractor --> Auditor
  Auditor -->|"Verify Claims"| SourceDocs
  SourceDocs --> Auditor
  Auditor -->|"Corrections"| Reconciler
  Extractor --> Pinecone
  Reconciler --> RiskReport
  Reconciler --> Flags

The Critique-Correct Loop

The key innovation is a dual-agent verification system:

  1. Extractor Agent: Parses ESG reports using Claude 3.5 Sonnet, extracting quantitative claims (emissions data, diversity metrics, supply chain audits)
  2. Auditor Agent: Cross-references every extracted claim against the original source PDFs, flagging hallucinations or unsupported assertions
  3. Reconciliation Engine: Resolves conflicts and produces a confidence-scored final output

This approach eliminates the hallucination problem that plagues single-agent extraction in high-stakes financial contexts.

Tech Stack

  • LangGraph — Multi-agent orchestration with conditional routing
  • Claude 3.5 Sonnet — Primary extraction and reasoning
  • Pinecone — Semantic search across historical ESG data
  • unstructured.io — PDF parsing and chunking

The Impact

MetricBeforeAfter
Report Processing Time3 weeks2 days
Analyst Hours per Report40+4
Greenwashing DetectionManual spot-checkSystematic
Risk Coverage$50M portfolio$4.5M flagged risk identified

The system now processes the firm’s entire portfolio quarterly, surfacing risks that would have gone undetected.