Talk to an engineer

Research

Research that ends in a running system.

A fixed share of revenue funds deep research every year, and it is not a side lab. We work on the problems our own engagements keep running into: models that have to be defensible, sensing at continental scale, control that cannot be paused, and data systems that stay honest under load. What the research produces ships, either into the products we license or into the systems a client runs.

Cabinets of a Cray supercomputer in a machine room

What has come out of it

Programmes.

Research here is judged by what it ends up running. These are the systems it has produced so far.

Land, water, air and carbon verification

BluVerify

Verify what is happening on the ground, backed by data. A map-based system that turns satellite and emissions data into dated, exportable assessments without needing a GIS specialist to drive it.

  • Fifteen environmental data products across optical, radar, thermal and air-quality sensors, with more than ten years of historical imagery behind every baseline.
  • Eighteen assessments covering drought, flood, landslide, wildfire, facility emissions and air quality, each returning a calibrated probability range rather than a single number.
  • Reports carry their maps, methodology and sources, and areas can be monitored with daily checks and automated alerts.
Who it is for
Government and disaster management, insurers and lenders, infrastructure and agriculture operators, and research institutions.
Built on
Copernicus Sentinel and Landsat imagery, Climate TRACE emissions data, and our own assessment models.
Visit www.bluverify.com

The retrieval, evaluation and guardrail layer

BluMind

The applied-AI research turned into a platform: retrieval that cites its source, agents with typed tools and approval gates, evaluation suites that gate a change, and a replayable trace behind every answer.

  • Permission-aware retrieval over a document estate, with answers bound to a citable passage and a designed refusal path when the evidence is thin.
  • Agent runtimes with durable state, bounded retries and a human approval gate on anything irreversible.
  • Golden sets, rubrics and continuous scoring of sampled production traffic, so quality regression is detected by the platform rather than reported by a user.
Who it is for
Any organisation putting a model into a workflow that has consequences if it is wrong.
Built on
Hosted and open-weight models, hybrid retrieval, OpenTelemetry tracing, and distillation onto hardware a client owns.
Inside BluMind

Research group

Applied AI

Making models useful where being wrong has a cost: grounding, perception, forecasting, and proof.

A GPU server blade laid out flat beside its cooling blocks and processors

Research group

Earth & Environment

Sensing the physical world from orbit and turning it into evidence somebody can act on.

Panorama of Srinagar with Dal Lake and the mountains beyond

Research group

Systems & Control

The layers underneath: autonomy on a plant floor, and data systems that stay correct at scale.

The switchyard of a 750 kV electrical substation

How it connects

Three groups, one practice.

The fields feed each other: perception needs evaluation, earth observation needs forecasting, and all of it needs data systems that stay correct.

01

Applied AI

Making models useful where being wrong has a cost: grounding, perception, forecasting, and proof.

See applied ai
02

Earth & Environment

Sensing the physical world from orbit and turning it into evidence somebody can act on.

See earth & environment
03

Systems & Control

The layers underneath: autonomy on a plant floor, and data systems that stay correct at scale.

See systems & control

Working on something at this edge?

If you have a problem that sits in one of these fields, or a dataset nobody has been able to make useful, write to us. An engineer who works on it will read it.

Talk to an engineer See our services