Leviathan — Sustainability & Anti-Greenwashing Audit Platform
Scikit-learn ensemble and Gemini reasoning to predict product carbon footprints and audit corporate greenwashing against ISO 14021.
Problem Statement
Consumers and enterprise procurement teams face pervasive 'greenwashing'—vague, unsubstantiated eco-friendly claims ('100% natural', 'carbon neutral') that mislead buyers and violate international advertising regulations.
Engineering Goals
- Train an accurate predictive model for product carbon footprint and durability using multi-modal product attributes.
- Implement a Greenwashing Inspector auditing marketing claims strictly against ISO 14021 environmental labeling criteria.
- Integrate external product knowledge bases via Open Food Facts and Google Shopping APIs.
- Deliver clear, auditable reasoning reports for procurement compliance.
Execution Pipeline & Mechanics
Product Ingestion
Pulls product composition, packaging materials, and manufacturer claims from Open Food Facts API.
Ensemble Footprint Modeling
Ensemble of Random Forest, Gradient Boosting, and Ridge Regression predicts lifecycle carbon emissions (R² = 0.88).
TF-IDF Vague Claim Extraction
Detects non-specific marketing buzzwords lacking measurable metrics.
ISO 14021 Regulatory Audit
Gemini reasoning maps specific claims against ISO 14021 clauses, classifying claims as Verified, Misleading, or Unsubstantiated.
Engineering Challenges & Design Decisions
Decision: Ensemble of tree-based models and Ridge regression over deep neural network
Rationale: Tabular product attribute datasets with high cardinality benefit from gradient boosted trees and regularized linear ensembles, providing both higher accuracy and interpretable feature importance.
Alternative Rejected: Deep MLP or TabNet architecture.
Core Architectural Highlights
Product metadata is fetched from public databases and shopping APIs. Numerical and categorical features are fed to a scikit-learn ensemble (Random Forest, Gradient Boosting, Ridge Regression). Textual marketing claims are extracted, vectorized via TF-IDF, and audited using Gemini reasoning prompted with ISO 14021 rules.
Testing & Empirical Evaluation
Cross-validated regression models against verified lifecycle analysis (LCA) environmental datasets.
Security & Reliability Defenses
Key Engineering Learnings
- •Combining quantitative tabular ensembles with LLM semantic reasoning provides far stronger regulatory audits than either technique alone.