D1
Instrument control & acquisition
Deterministic control loops, timing-critical acquisition, and telemetry pipelines that do not lose samples when the network hiccups. Provenance recorded at the point of capture.
04 The lab
Software with the rigour of the science it serves.
Research software fails differently from business software. It fails quietly, in the fifth decimal place, eighteen months after the result was published. We engineer against that specific risk.
01 Domains
Where we are useful to a research programme.
We are engineers, not domain scientists. The division of labour is clear: you own the science, we own whether the software is worthy of it.
D2
Numerical & simulation workloads
Solvers and simulation pipelines taken from a single workstation to a scheduled, reproducible, cost-bounded cluster workload — without a rewrite that the scientists cannot read.
D3
Signal, image & spectral processing
High-volume processing pipelines with validated numerics: calibration, registration, denoising, feature extraction and downstream inference, versioned end to end.
D4
Research data platforms
Catalogued, queryable, citable data with lineage from raw capture to published figure. Storage tiering designed around access patterns rather than convenience.
D5
HPC & GPU orchestration
Scheduling, queue strategy, containerisation and cost control across on-premise clusters and burst-to-cloud capacity. Utilisation reported, not assumed.
D6
Scientific software hardening
The grant-funded codebase that now has forty users. Tests, packaging, CI, numerical regression suites and documentation — without breaking the science it encodes.
02 Standard of rigour
The practices that keep a result defensible.
Everything here exists because we have seen its absence cost someone a year.
- 01
Reproducibility as a build artefact
Pinned environments, containerised toolchains and deterministic seeds. A result from eighteen months ago can be regenerated bit-for-bit, or the system tells you why not.
- 02
Numerical regression testing
Golden datasets with tolerance bands, so a refactor that shifts the fifth decimal place is caught by a pipeline rather than by a reviewer six months later.
- 03
Provenance and lineage
Every derived dataset carries its inputs, code version, parameters and instrument state. Auditable by a third party without archaeology.
- 04
Uncertainty carried through
Error propagation treated as a first-class concern rather than dropped at the first aggregation. Confidence reported wherever a number is displayed.
- 05
Performance with a budget
Profiling before optimising, and a stated cost per run. Compute spend is a scientific constraint, not an infrastructure detail.
- 06
Software the group can own
Written to be read by domain scientists. Reviewed with them, documented for them, and handed over with the intent that we become unnecessary.
03 Collaboration models
Four ways we plug in.
Research funding and procurement rarely fit a standard consulting shape. These are the arrangements that have worked.
C1
Research group partnership
Embedded engineering capacity alongside a research team — building the platform, pipeline or instrument software while the group focuses on the science.
C2
Instrument & hardware builders
The software layer for a physical instrument: firmware interfaces, control software, acquisition, analysis and the customer-facing application.
C3
Grant & consortium work
Named technical partner on funded programmes, including deliverable definition, work-package ownership and the reporting that comes with it.
C4
Internal R&D for industry
Corporate research teams who need a prototype taken to the standard where the rest of the business can adopt it.
04 Open notes
We publish the method.
Not gated, not summarised for a buyer. Written for the engineer who has to make it work.
- Field note Measuring an AI-native delivery uplift honestly Everyone reports a productivity gain from AI-assisted engineering. Very few can show a baseline. Here is the instrumentation we insist on before claiming a number. AI-native delivery Engineering leadership Measurement 5 min
- Method The evaluation harness is the product In industrial AI, the model is the cheap part. What determines whether a system survives its first bad week is the apparatus you built to measure it. Applied AI Manufacturing MLOps 6 min
- Essay What a digital twin owes you Most things sold as digital twins are dashboards with better rendering. A useful test: can it be wrong in a way you would notice? Manufacturing Simulation Digital transformation 5 min
Research brief
Bring the instrument, the dataset, or the code you inherited.
A short technical conversation about what the software has to guarantee, and what it currently does not. Useful whether you are scoping a grant deliverable, hardening an existing codebase, or building an instrument from scratch.
or write directly — [email protected]