Project
Internal Fact Checker
Editorial AI tool grounding article claims in live evidence using FastAPI, LangGraph, and Gemini search.
- 50+ editorial staff
- Grounded search RAG
- Active review loop
FastAPILangGraphGeminiVertex AIpgvectorPython
NerdWallet’s brand is built on financial accuracy and editorial trust. Before publishing, financial content must be verified against current interest rates, regulatory changes, and institutional policy updates. The Internal Fact Checker accelerates this editorial process without sacrificing editorial rigor.
The System Design
Built in collaboration with NerdWallet’s Editorial leadership, the Fact Checker runs as a specialized microservice:
- Claim Extraction Pipeline: FastAPI and LangGraph workflows parse draft articles and isolate discrete factual and numerical assertions (such as APR ranges, fee thresholds, and bonus terms).
- Grounded Evidence Retrieval: Claims are queried against live web evidence and internal knowledge bases using Google Vertex AI Gemini with Search Grounding and pgvector vector storage.
- Structured Verdicts & Rewrites: The engine returns structured evaluations classifying claims as Verified, Outdated, or Unsubstantiated, supplying direct citation links and proposed editorial adjustments.
- Editorial Feedback Loop: Interactive feedback interfaces allow editors to flag false positives and false negatives, continually refining the prompt evaluation boundaries.
Outcome
V1.1 was successfully rolled out to over 50 editorial team members, substantially reducing manual research turnaround time while safeguarding NerdWallet’s high editorial standards.