Aletheai · founder and CEO since December 2024

Decisions with the evidence attached.

Aletheai is a behavioural decision-intelligence company. Its engine reads behavioural signals from the systems a business already runs, ranks who or what deserves attention by the evidence, attaches a plain-English reason and next action to every item, and declines to rank where the evidence is insufficient.

First market: UK estate agency, launching through the Alto (Zoopla Group) marketplace; first cohort recruiting.

Original work

Six patent applications. One principle.

Six patent applications filed, 2025–2026 (UK and international). Two are summarised below; the others are unpublished. A machine decision is only worth the hours it redirects, and it is only followed if it is honest about where it cannot see.

Patent application · international (PCT), priority September 2025

Behavioural decision intelligence

An engine that ranks who or what deserves attention by the evidence rather than a score over everything, attaches a plain-English reason and next action to every item, and refuses to rank where the evidence is insufficient, with the refusal recorded. Its behavioural archetypes live in a fixed, versioned library that changes only at signed releases, never silently in production.

Patent application · UK, September 2026

Action gating for AI agents

Every act an AI agent takes passes through a gate the agent cannot reach, is registered before it happens, and must be justified by evidence the agent did not write. The gate uses no model: it is deterministic and replayable, so oversight does not depend on reading the agent's reasoning. Governance of machine conduct, not machine thought.

Research questions

What the company is trying to answer.

When should a machine decline to answer?

Ranking only where the evidence supports it, and recording the refusal, so a person can see where the system could not see.

How do you govern what an AI agent does without reading its mind?

Gate the action, not the reasoning: evidence the agent did not write, checked by a gate it cannot reach.

What does it take for a decision to be defensible afterwards?

A record a person can open: what was known, what was inferred, and why this item came first.

Research partners and recognition

Partners and recognition.

STFC Hartree CentreFeasibility study on real human-behaviour datasets, including how to treat missing data before trusting a model on it.
RTC NorthBusiness support.
AI ForgeFinalist, 2025.
Read more

The Astra Moment

The most powerful AI yet just moved part of how it thinks into a form no human can read. That makes the case, urgently, for the other kind of AI: the kind that keeps a ledger you can open.