GreenGuard M&A: AI Sustainability Public Company Deal-Screener Prototype for Investment Banking

In investment banking, traditional financial due diligence frequently misses hidden compliance liabilities. This project creates an automated screening framework designed specifically for a solo financial analyst to deploy. (fully written by myself) Business Requirements (Version 1). We will build a multi-agent AI software to detect potential green-washing risks of public companies in the EU. The tool will use probabilistic LLM and deterministic tools to vectorize annual reports in XBRI format. The tool will extract relevant quantitative KPI metrics (% of Taxonomy-eligible and % of Taxonomy-aligned turnover, CAPEX, and OPEX) in a deterministic way and qualitative data in a probabilistic way from the vector database based on similarity to the following query: cross-check the % of taxonomy-aligned and the % of taxonomy-eligible turnover, CAPEX, and OPEX, with sustainability claims in the annual report to highlight potential greenwashing risk of the company on a continuum score from 0-100 based. (AI-written) Because you are working entirely solo, the project relies strictly on public-domain financial data, open-source data science libraries, and existing regulatory disclosure databases. The final deliverable will be a private GitHub repository containing the open-source code. This case study will explicitly demonstrate how the tool identifies material risk mismatches, creating a highly visible, elite proof-of-concept that showcases your practical expertise directly to institutional hiring managers. (fully written by myself) Business Requirements (Version 2): A private, zero-cost, locally hosted, enterprise-grade (reliable) multi-agent AI software that can decide whether a given public company is showing proof of potential greenwashing or not on a continuum greenwashing risk score of 0-100 through advanced analysis of the corporate's XHTML annual report by extracting the taxonomy KPIs from the report in a deterministic way, vectorizing the annual report and extracting Sustainability-related statements from the vector database using semantic search, and separately vectorizing the EU Taxonomy, CSRD, the ESR, the Disclosures Delegated Act, Climate Delegated Act, and the Environmental Delegated Act, and extracting the relevant sustainability rules using semantic search, and 1) compute the difference of the embeddings, if any, between the quantitative sustainability KPIs and the qualitative sustainability-related statements of the annual report and 2) compute the difference of the embeddings between the sustainability-related embeddings of the combined regulatory vector database, and the sustainability-related embeddings of the annual report's vector database, and 3), deterministically extract the NACE code or codes of the target public company from its annual report, and find which embeddings from the combined regulatory vector database apply to it, and use that subset of embeddings exclusively during the embedding distance computation between the regulatory vector database and the corporate vector database. (fully written by myself) Business Requirements (Version 3) (Highly Technical): 1. Generate embeddings for the corporate data and the regulatory data. 2. Run semantic search with q1 against the corporate vector database, where q1 = 'which investments into economic activities does the company claim to be environmentally sustainable, if any?'. Store the results in a new vector database named company_reported_sustainable_investments. 3. Run semantic search with q2 against the regulatory vector database, where q2 = 'which articles, paragraphs, subparagraphs, and technical standards apply to company_reported_sustainable_investments ?'. Store the results in a new vector database named applicable_regulatory_regulations_and_technical_standards. 4. Compute the distance between the vector databases of company_reported_sustainable_investments and applicable_regulatory_regulations_and_technical_standards. Store the result as a variable named computed_distance_between_corporate_data_and_regulatory_rules 5. Deterministically extract the Taxonomy KPIs % of Taxonomy-eligible, Taxonomy-aligned, Turnover, CAPEX, and OPEX from the company's XHTML file. 6. Run semantic search with q3 against the corporate vector database, where q3 = 'Are the sustainability statements backed up by the Taxonomy KPIs?". Store the results as sustainable_goals_validity. 7. Compute a potential green-washing risk score between 0-100 based on the computed_distance_between_corporate_data_and_regulatory_rules variable. 8. Formulate an Executive Assessment based on the computed_distance_between_corporate_data_and_regulatory_rules variable and the sustainable_goals_validity. Clarify any potential material discrepancy using 'environmentally sustainable investments' as the subject.