FIFA Brand Compliance Review System

Multimodal RAG compliance review system for Lenovo FIFA marketing materials.

This project was built during my AI engineering internship at Lenovo (Beijing). The goal was to automate pre-release compliance checks for co-branded FIFA marketing assets and replace slow, inconsistent manual review.

The system covered four linked stages:

  • Parse Lenovo Brand World pages and FIFA branding PDFs into structured, traceable rules.
  • Index those rules with hybrid retrieval over vector and keyword signals.
  • Detect logos, spacing, proportions, colors, and text usage from submitted images.
  • Produce rule-grounded compliance judgments with evidence and remediation suggestions.

The implementation stack included LangChain, Milvus, BGE-Reranker, PyMuPDF, PaddleOCR, FastAPI, and multimodal vision models. Structured outputs were validated with Pydantic, and every decision was linked back to source rule IDs, page numbers, and annotated evidence.

Measured outcomes from the project:

  • Recall@10 improved from 76.8% to 92.4% after hybrid retrieval and reranking.
  • Faithfulness increased to 95.1% for rule-grounded answers.
  • Single-image review P50 latency dropped from 74.2s to 32.8s with caching, dynamic top-k compression, and parallel orchestration.

This work pushed me to treat RAG not as a demo pattern, but as an evidence system where retrieval quality, visual measurement, and rule traceability all matter to the final answer.