MergeSeat MergeSeat
AI systems engineering · Netherlands

Your data, your model, your decision.

MergeSeat builds AI systems that make the people who use them independent. We build the system for your task, run it where you say, with the model you choose, and hand it over with its error rates written down.

What an engagement looks like See what we have built

The concept

Independence, not a subscription.

The industry sells AI as a subscription to someone else's black box. MergeSeat sells the way out.

Most of any pipeline is mechanical, so it is repeatable, cheap, and inspectable before any model is paid for. A person holds the merge seat, so the machine's answer can be checked, refused, and signed for.

The name

The merge seat is the chair where a person decides whether a machine's proposal is accepted. Agents propose, a human merges. The company is named after that decision.

↑ Back to top

Why we can be believed

Seven things that are true today.

Open by commitment

You can read the code you are trusting.

Axial, diligence-reader, Neo4All, G-Lab, AEO, SocioRAG and SpecClass are public repositories.

Runs where you say

Your data stays on your premises or in your account.

CIP is self-hosted by design. Axial runs on local corpora.

Mechanical where it can be

A model is used only where it earns its cost.

Axial chunks deterministically, model-free, before any inference is paid for.

Modular by design

Every method leaves as a package or an MCP server.

AEO ships as a plugin: 15 skills, five roles, five gate scripts.

Measured, not demonstrated

Tools ship with their limits written down.

Axial: validated across about 30 sources and 17k chunks, two gates still open by design. diligence-reader: 100 on a public rubric, twice, spread zero, every dollar in a ledger the code wrote.

A person holds the merge seat

Approval and refusal are enforced by gates, not prompts.

AEO hard-blocks the merge. A hook returning exit code 2 is a constraint.

Proven on hard material

Built where a wrong answer costs more than a slow one.

CIP reads Arabic, Persian, Hebrew and English conflict media and tests it against measured behaviour.

↑ Back to top

The products

The same idea, built in a different room.

Each one is proof of the concept rather than an item in a catalogue.

Flagship

CIP

Conflict Intelligence Platform

An intelligence agency in a box.

It reads Arabic, Persian, Hebrew and English media across a conflict theatre into actors, events and relationships, then tests that picture against measured behaviour: strait transits, airspace density, economic indicators. Text is what parties assert. Movement and money are what they do.

Every claim traces to its source. First theatre: Iran and the Gulf.

Production tested Proprietary, closed source Self-hosted

diligence-reader

Public, open source · Apache 2.0

Reads an acquisition data room, the hundred-odd contracts, financials, minutes, emails and spreadsheets a buyer's team gets before a deal closes, and writes the findings report: a recommendation with a number, findings ranked by money at stake, the material matter quantified with a deal action, the lesser issues, the open items. Every sentence cites the page it came from, and every citation, figure and certainty word is checked against that page by code before the report is allowed out.

Seven stages, two of them model calls, the rest deterministic code that gives the same answer twice. On a public synthetic room with an answer key it scores 100 on the rubric, both runs, at about $0.30 of model cost. On the real Verizon and Yahoo filings it lands on the price cut and the liability split. A fixed rebuild of John Adeojo's recursive language model run; credit in the repository.

Axial

Public, open source

Turns a corpus of academic books into original comparative-historical analysis. Every claim is marked for what kind of claim it is, points at the passages that ground it, and carries a disclosed confidence band. Chunking is deterministic and model-free, inspectable before any spend, and the release gates sit outside the model's control.

Phase A pipeline complete end to end, validated across ~30 sources and ~17k chunks. Two gates still open by design: human labelling of the gold set, and full-corpus runs held until evaluation closes. Domain logic lives in versioned schemas, so porting to another country is a schema edit, not a code change.

Neo4All

Public, open source

Turns structured and unstructured documents into governed Neo4j knowledge graphs for investigative work, so a finding can be defended line by line rather than asserted. Proposal, approval, diff, execution, audit: nothing changes the graph without a recorded decision.

G-Lab

Public, open source

The companion workbench: builds and interrogates those graphs to surface conflict-of-interest patterns, financial flows and cross-border dynamics for researchers and journalists. Self-hosted, read-only against the graph, no telemetry.

AEO

Public, open source · the method, not for sale

A Claude Code plugin: 15 skills across five agent roles, governed by five gate scripts. It automates planning, triage, building and review, and hard-blocks the merge so only a person can ship. It is how MergeSeat itself is built.

SocioRAG

Public, open source

Socially-grounded document retrieval, shipped as a package that leaves with the client.

SpecClass

Public, open source

Document classification with a confidence score per result, shipped as a package that leaves with the client.

flood-fire

Public, open source

Geospatial flood and fire detection with validation gates on every result.

↑ Back to top

Writing

The Merge Seat

The AI proposes, the human disposes, and something mechanical stands between them. Notes on building systems that work that way.

Read the publication
↑ Back to top