Four letters. Four million times.
A bacterial genome is almost entirely context. A sequencing run returns all of it and ranks none of it.
Escherichia coli reference genome: 4,641,652 bpEnterprise AI
We build enterprise AI.
The models, the platforms they run on, and where it matters the instruments underneath them. Four of our own technologies are in deployment, and the same engineers build for clients.
Antimicrobial resistance
A gene can be present and still be doing nothing at all.
Which is why finding one is not the same as explaining why a drug has stopped working.
Sequencing gives you everything an organism is carrying. Millions of letters, all of it at once, none of it ranked. The usual approach is to compare that against a list of the resistance genes somebody has already described. It is fast, and for organisms that have been studied to death it works well.
The part that matters is very small.
A resistance determinant is often fewer than a thousand bases. Finding it is a search problem. That was solved years ago.
blaNDM-1 coding sequence: 813 bpPresence is not resistance.
A gene can sit silent. Truncated. On a plasmid the organism no longer maintains. Presence and phenotype disagree often enough that presence alone cannot be reported as a result.
This is where a catalogue stops.
We name the protein.
The model reads context rather than identity. It returns the protein doing the work, and the drug class it defeats.
Output: the protein, and the drug class it defeatsWhat we built
So we taught it to read instead.
It returns the protein doing the work, and the drug class it defeats.
Not a database match. An attribution, including for proteins that are not in anyone's catalogue yet.
It arrives in time to matter
Culture takes days. Our call lands while empiric therapy is still being chosen, which is the moment a broad-spectrum drug usually gets reached for out of caution.
It runs where the samples are
On-premise, in-country, on hardware a laboratory already owns. Patient data does not have to travel for the answer to arrive.
It is tested on your organisms
Send isolates you already have phenotypic results for and compare. That is a better test than anything we could write here.
Our own technology
Four things we built because nobody else had.
Each started here as a research problem with no customer attached. Each is now a system in production.
Genomics
AMR protein detection
The protein defeating the drug, read out of sequencing data.
Read more →PredictionEarly cancer risk prediction
Cervical and breast cancer risk, up to three years before it presents. A reason to look sooner, not a diagnosis.
Read more →ImagingRadiology AI
Cancer detection from diagnostic imaging, reported at 96.4% accuracy against what the radiologists agreed on.
Read more →DentalAligner treatment planning
Three to four weeks of orthodontic planning, returned in a day.
Read more →Planning
A day instead of a month.
And almost none of that month was ever computation.
1dayFrom intraoral scan to a plan a laboratory can manufacture.
It was queueing. A case waits for a technician. The technician produces a setup. The clinician reviews it and sends it back. It waits again. Each round trip is a day of work and about a week of calendar. So we put the clinician inside the pipeline rather than at the end of it.
Compute
We use GPUs. Our systems do not depend on them.
Scriptics was training models to run on ordinary processors before accelerated compute became the default in this industry.
That discipline stayed. We use accelerators where a workload earns them. What we design out is the assumption, because a system that assumes an accelerator fleet can only be installed where one has been bought. Ours run on hardware the customer already owns, which is how a diagnostic ends up working in a hospital basement rather than only in a cloud region.
Services
Six practice areas.
Around thirty capabilities, delivered by the engineers and scientists who write the code rather than by a layer above them.
01
AI and automation
Agentic systems, machine learning, language and vision, and the evaluation work that decides whether any of it is safe to ship.
Read more →02Engineering and cloud
Modernisation, data platforms, DevOps, cybersecurity, embedded systems and cloud rendering.
Read more →03Product and design
Full-stack engineering, interface design, and proofs of concept built to answer one question.
Read more →04Data and forensics
Data engineering, digital forensics, scientific computing and optimisation.
Read more →05Geospatial engineering
Satellite and radar imagery, weather and climate data, spatial analytics and earth observation platforms.
Read more →06Blockchain and Web3
Smart contracts, asset tokenisation, permissioned networks and digital identity.
Read more →Industries
Nine sectors, one standard of proof.
The regulatory bar moves between them. The way we validate does not.
01
Financial services
Risk, fraud and decisioning systems under audit, and the platforms that carry them.
02
Healthcare
Clinical and diagnostic systems, built to an intended use and through validation.
03
Manufacturing
Visual inspection, process optimisation and the instruments on the line.
04
Gaming
Intelligent behaviour, procedural content and real-time player analytics.
05
Logistics
Routing, scheduling and resource allocation as optimisation problems rather than dashboards.
06
Cybersecurity
Detection, response and digital forensics across the estate.
07
Energy
Weather and climate pipelines, forecasting and asset intelligence.
08
Retail
Demand, personalisation and the data platforms underneath them.
09
Insurance
Climate and catastrophe modelling, claims automation and portfolio analytics.
How we work
Tell us what has to be true.
The budget, the compute, the regulation, the date. The design follows from those.
And if it does not close, you will hear that before the contract rather than at the second milestone. Most of our work comes from clients who came back, and a project that should never have started ends that.