Skip to content

Overview

Ariadne is a sensemaking harness for entities that live inside large organizations.

The problem

Intelligence about a single entity is distributed across systems that were never designed to interoperate:

  • graph databases for relationships, hierarchy, and networks
  • relational stores for records, attributes, and facts
  • unstructured repositories for free text, documents, and transcripts

The content spans modalities too: metadata, text, imagery, and video. The barrier isn't getting into any one store; it's reasoning across all of them at once. The evidence linking two facts often exists only through an implicit organizational relationship, sitting in a different store and format than either fact.

The approach

Ariadne uses the Claude Agent SDK as a single analytic interface over these systems. An orchestration layer dispatches specialized tools to retrieve, interpret, and synthesize evidence across graph, relational, and unstructured sources, without replacing the infrastructure underneath.

The research question: what tools, skills, and hooks does a harness need to support a rigorous end-to-end analytic workflow over entities in an organizational hierarchy? Ariadne prototypes the minimum viable set: database connectors, modality processors (image and video analyzers, text extractors), and hierarchical reasoning hooks.

The deliverable

A working prototype that runs an end-to-end workflow. It takes a target entity or organizational node as input, runs a coordinated sequence of tool calls, and synthesizes the evidence into a cited analytic product.

It is judged on four things:

  1. Traverse organizational relationships.
  2. Reconcile evidence across modalities.
  3. Reduce the analyst's manual-pivot burden.
  4. Surface non-obvious connections that conventional tooling would miss.

The umbrella role

Ariadne is an umbrella effort. Rather than duplicate work, it defines integration interfaces so contributions from sibling projects (graph-extraction pipelines, entity-resolution models, multimodal indexing) surface as callable tools inside the harness. That makes it both a standalone research contribution and a demonstration layer for the wider portfolio.

Where to next