We're a group of Stanford students building Subline: a browser-based tool that surfaces a knowledge graph extracted from newsroom archives. It identifies entities (people, places, laws, organizations) and maps their relationships (WORKS_FOR, VETOED, SUPPORTED_BY, etc.) across time. This turns static archives into a living, queryable network. It builds a system that preserves context and makes it accessible to new journalists through intuitive prompts. Ask a question like: "How is Jacob Frey connected to water security issues via corporations that supported his mayoral opposition?" —and you get back not just text, but the precise, source-backed relationships that answer it. Every node is linked to the original reporting. Every path tells a story.

This isn't just search—it's investigative inference. It helps junior reporters grasp decades of background in minutes. It flags contradictions: a politician's public stance on housing policy vs. their voting record. It surfaces overlooked connections that might spark new leads. And crucially, it's built with journalists, for journalists.

Subline is an AI-native data extraction tool. We use Natural Language Processing (NLP) in combination with AI to extract our knowledge graph. Then, we use machine learning to surface the right information to users.

One of the core challenges of this project is ensuring the accuracy and trustworthiness of the information that Subline surfaces. In journalism, even minor inaccuracies can have significant consequences, so we're taking rigorous steps to put humans in the loop and ensure data integrity.

• We're training our own NER model
• We've got "human in the loop" verification at each extraction step
• All extracted information is dated and cited

Another important challenge is creating seamless integration into real newsroom workflows. If the tool is overly complex or time-consuming, it won't be used. That's why we're working closely with journalists at the Minnesota Star Tribune, focusing on intuitive, low-friction interfaces that provide immediate value in the reporting process.

One example of that is our natural language to cypher query feature that translates questions to precise cypher queries, so that users can take advantage of the powerful cypher query framework without the syntactic baggage.