KEYSTONE: Building an Evidence-First AI Company Brain with FalkorDB

revanth2444-cmyk — Sun, 18 Oct 2026 09:49:31 +0000

The Problem: Company Knowledge Is Connected, But Often Scattered

Companies create information every day: product plans, project updates, decisions, meeting notes, and risk reports. The challenge is not just finding information — it is understanding how different pieces of information connect.

Imagine a product launch plan that contains two different dates: December 2026 in one claim and January 2027 in another.

A basic search might return both statements without explaining their implications. An AI assistant might even present one date as the answer without enough evidence.

I wanted to explore a better approach.

Introducing KEYSTONE

KEYSTONE is an Evidence-First AI Company Brain built with Python, Streamlit, and FalkorDB.

Its goal is to help users explore connected company knowledge, identify potentially conflicting claims, and trace information to related projects, owners, and risks.

The principle behind KEYSTONE is simple:

AI should not just answer a question. It should show the evidence that helps people understand the answer.

Why FalkorDB?

FalkorDB is the foundation of KEYSTONE's knowledge graph.

Instead of treating every fact as an isolated record, KEYSTONE represents company information through entities and relationships. Documents can support claims, people can own projects, and projects can be connected to risks and decisions.

These relationships allow the application to retrieve connected evidence and help users investigate a business situation.

The graph is not just a visualization. It plays a central role in how the prototype organizes and retrieves company knowledge.

What KEYSTONE Can Do

1. Visualize the Knowledge Overview

The dashboard displays connected company information, including documents, projects, claims, people, decisions, and risks.

2. Detect Potential Conflicts

In my demonstration, the Product Launch Plan contains two claims:

claim001: The product launch is planned for December 2026.
claim002: The product launch is planned for January 2027.

KEYSTONE flags these claims as a potential disagreement and encourages the user to review the underlying evidence.

It does not automatically assume that either date is correct. Detecting a conflict and resolving it are different tasks.

3. Connect Projects, Owners, and Risks

The prototype connects project information with ownership and risk.

For example, the Product Launch project is shown as at risk, with an identified project owner and conflicting launch dates associated with the risk.

This helps users move beyond an isolated statement and investigate the wider business context.

4. Answer Questions Using Connected Evidence

Users can ask questions about company knowledge and inspect the supporting claims retrieved by KEYSTONE.

When asked about the launch, the prototype returns both conflicting dates and their source document instead of pretending that the evidence establishes one definitive answer.

What I Learned

Relationships Matter

Facts become more useful when they are connected. A claim about a launch date becomes more meaningful when linked to the affected project, its owner, and the relevant risk.

Evidence Matters More Than Confidence

A confident-sounding answer is not necessarily a reliable answer. I learned the importance of making uncertainty visible and allowing users to inspect the evidence behind a conclusion.

A Graph Database Must Do Real Work

Using FalkorDB challenged me to think about how graph relationships contribute to the application's actual behavior, rather than using a graph only as a visual feature.

Deployment Is Part of Building

Connecting to FalkorDB Cloud and deploying the Streamlit application taught me about database configuration, environment variables, credentials, and making a prototype accessible online.

Limitations Should Be Clear

KEYSTONE identifies potential disagreements, but a flagged conflict does not automatically establish which statement is correct. Source review and additional context may still be needed.

What Comes Next?

KEYSTONE is a prototype, and there is more to explore. Future improvements could include richer document ingestion, temporal tracking of claims, more advanced conflict resolution, and deeper tracing of how new information affects previous decisions.

My long-term goal is to make company knowledge more connected, traceable, and useful for decision-making.

Final Thoughts

Building KEYSTONE helped me explore how graph databases can support more evidence-aware AI applications.

The project is built around a straightforward idea: when people make decisions, they should be able to investigate the evidence and relationships behind the information they receive.

KEYSTONE — Connect the evidence. Understand the context. Make better-informed decisions.

Built for Graph Hacks: Context for AI Agents, powered by FalkorDB.

Project repository: https://github.com/revanth2444-cmyk/keystone-graph

Demo video: https://youtu.be/JRdFwYEhxNA

DEV Community: revanth2444-cmyk