Skip to content

, Industries

Domains where the cost of being wrong is legible.

We work in a small number of sectors on purpose. Each one has a person who can tell you whether the output was right and a reason they care, which is the precondition for building anything measurable.

01How we choose

We avoid sectors where nobody can tell whether the answer was any good.

It is a genuine filter and it costs us work. If a use case has no observable ground truth, no reviewer, no downstream consequence, no way to check, then there is no evaluation set to build, and without one we are just shipping a demo with a support contract attached.

The sectors below share three properties: the output is checked by somebody qualified, being wrong has a named cost, and the domain has enough structure that correctness can be specified. Those are the conditions under which the way we work is worth paying for.

Where we have not worked in a sector before, we say so, and we price the discovery honestly rather than claim depth we do not have. Experienced engineers, no invented history, that is a position we intend to still be holding in twenty years.

02Sectors

Six domains, and what actually makes each hard.

01

Healthcare & life sciences

The reality

Clinical AI is being adopted faster than the review capacity to supervise it. The dangerous output is not the obviously wrong one, it is the fluent, well-formatted note that quietly contains a claim the encounter does not support.

Where we work

  • Clinical documentation with claim-level grounding
  • Coding and prior-authorisation support
  • Literature and protocol retrieval with provenance
  • De-identification pipelines and audit trails

The non-negotiable

Every generated statement traceable to a source in the record, and an explicit surface for the ones that are not.

02

Financial services & insurance

The reality

There is an auditor at the end of every decision path. A model that improves a metric but cannot explain a single decision is not deployable, however good the metric.

Where we work

  • Underwriting and credit decision support
  • KYC, onboarding and document verification
  • Surveillance, AML alert triage and narrative drafting
  • Regulatory reporting and control testing

The non-negotiable

Reproducible decisions, versioned models and prompts, and a record of what evidence produced which output.

03

Manufacturing & mobility

The reality

The useful work happens where the connectivity is worst and the hardware is oldest. Cloud-first architectures quietly assume conditions the shop floor does not have.

Where we work

  • Visual inspection and defect classification at the edge
  • Maintenance intelligence over service history
  • Field data capture that degrades predictably offline
  • Telemetry and diagnostics for connected vehicles

The non-negotiable

A stated accuracy budget on the hardware that exists, and defined behaviour when the network is gone.

04

Energy, climate & ESG

The reality

Disclosure regimes have converted sustainability reporting into an assurance problem. The output is a number somebody has to certify, and estimation methodology is now audited.

Where we work

  • Emissions and disclosure data pipelines
  • Supply-chain document extraction at scale
  • Methodology-transparent estimation with uncertainty
  • Assurance-ready evidence packs

The non-negotiable

Every figure decomposable to its inputs, its method and its assumption set.

05

Legal & compliance

The reality

The value is in the citation, not the summary. A confident paragraph with no provenance creates work for a reviewer rather than removing it.

Where we work

  • Obligation extraction as structured, citable objects
  • Contract review and deviation flagging against a playbook
  • Regulatory change monitoring and impact mapping
  • Policy and precedent retrieval with version awareness

The non-negotiable

Output that is queryable and traceable to a clause, not prose that has to be re-verified from scratch.

06

Public sector & infrastructure

The reality

Citizen-facing systems have to work in several languages, for people with no ability to detect a wrong answer, and be inspectable by parties who are not obliged to trust the vendor.

Where we work

  • Multilingual and code-switched service interfaces
  • Scheme eligibility and entitlement guidance
  • Document processing at population scale
  • Grievance triage and routing

The non-negotiable

Measured performance per language, published refusal behaviour, and no silent degradation for the languages with the least test data.

Not on this list?

Tell us the problem anyway. If the answer can be checked by somebody who knows the domain, it is probably work we want.

Start a conversation