Services

We build document and knowledge-graph pipelines for teams whose data cannot leave their environment. We use the open-source tools we maintain, Verdant and GraphForge, and you keep everything we build.

Start with a pilot

A pilot is a short, fixed-scope project on one set of your documents.

1. Extract. We test extraction options on your documents with Verdant and score each against samples you approve.
2. Connect. We pull out the entities and relationships that matter and load them into GraphForge.
3. Ask. We answer a handful of your real questions from the graph, with the evidence attached, and draw the results where a picture helps.

You get the pipeline, the code, an evaluation report on your own pages, and a plain recommendation on whether to go further. We say so when a different tool is the better fit.

Talk to us and bring ten documents that give you trouble.

Who we build for

Data and AI teams

Your retrieval or agent pilot works on clean text and fails on the documents that matter: scans, tables, charts, and long agreements. Sending those documents to an outside service is not an option.

• A pipeline in your environment. It runs on your infrastructure with the model you already have access to, including models behind Amazon Bedrock or a private endpoint.
• Measured quality. We score the output against samples you approve, so you see accuracy on your own pages before you commit.
• Access for people and agents. Your team gets a web UI and a CLI. Your agents connect through MCP.
• The right parser for the job. We use Verdant to compare parsers and models on your documents and keep the one that does best for the cost.

Investigations and due diligence

A case arrives as thousands of documents and one question: who is connected to whom, and how do we know? The work has to be defensible, and the material cannot leave your machines.

• A graph of the case. People, organizations, events, and documents, connected and queryable in GraphForge.
• Evidence with each claim. Each assertion carries the evidence behind it and a confidence assessment, so a reviewer can check the reasoning.
• Questions in plain language. Analysts work through an agent and do not need to write queries.
• Nothing to provision. GraphForge is embedded. A case can live on one analyst's laptop.

GraphForge is pre-1.0. We scope each engagement around what it does reliably today.

Agent and retrieval builders

Your agent needs to remember entities and relationships, or your retrieval system needs to follow connections between documents, inside your application rather than as another service to operate.

• An embedded graph. GraphForge runs in your Python or Node process, with openCypher queries, graph algorithms, and search in one engine.
• Integration with your stack. We connect GraphForge to the retrieval or agent framework you already use.
• A license you can ship. GraphForge is Apache-2.0.

For a centrally hosted application with many concurrent users, an operational graph database is the better fit. We will say so.

Research and heritage

Research and heritage work starts from sources that are hard to use: scanned records, handwritten registers, and collections described in many formats.

• Use the tools free. Verdant and GraphForge are open source.
• Publish a graph. GraphForge Hub gives public graph repositories a stable address others can browse and clone.
• Work with us on a grant. We join funded projects as a technical partner for extraction, graph modeling, and training. Our background includes community-history and genealogy graphs and extraction to the CIDOC-CRM heritage model.

After the pilot

• Implementation. We extend the pilot pipeline to the full document set and the questions your team actually asks.
• Adoption help. Training for analysts working through an agent, and for engineers integrating GraphForge or Verdant.
• Support and subcontracting. Ongoing support for teams running the stack, and subcontracting for consultancies that hold the client relationship and want document and graph depth.

Next steps

• Talk to us about a document set, a case, or an integration
• Try the tools first: open source