AI × ESG: Generative Scoring for Circular Fashion (GreenSync)

Project Overview

Fashion ESG evaluation is often fragmented and qualitative. GreenSync builds a lightweight, reproducible pipeline that turns images and documents into structured ESG signals for circular fashion. It demonstrates: (E) image understanding for material/impact cues, (S) policy/document assessment via prompting, and (G) structured JSON scoring with schema validation.

What I Built

• Designed an end-to-end pipeline: image encoder → document QA → rubric-based scoring → JSON aggregation.
• Created a scoring schema (E/S/G dimensions, weights, rationale fields) and enforced it with JSON Schema validation.
• Built deterministic prompt blocks and fallback rules to reduce hallucinations and keep outputs schema-compliant.
• Implemented material tag extraction from product images and policy checks from brand documents/FAQs.
• Evaluated with hand-labeled samples: precision/recall on environmental tags; sanity checks on policy claims.
• Packaged as a Kaggle Notebook with pinned environment and runnable examples for transparent reproducibility.

Reflection

Multimodal ESG scoring works only with strong guardrails. The key was agreeing on a strict JSON schema, writing assertive prompts, and validating every output. Future iterations: enlarge labeled sets, add confidence scores, human-in-the-loop review, and task-specific fine-tuning or adapters to improve robustness.