Forty million papers. One prescription.

The problem

Genomics arrived. The last mile did not.

The science is settled and the tests are being done. What fails is the end of it — where a result has to become a decision, made by one person, quickly, with no way to check what the literature says.

£2.21bn

Projected annual cost to NHS England of admissions caused by adverse drug reactions.

Osanlou et al., BMJ Open, 2022

2 in 5

Of those admissions were judged avoidable.

Osanlou et al., BMJ Open, 2022

1 in 7

Acute inpatients carry a gene–drug interaction that could change what they are prescribed.

McDermott et al., QJM, 2025

30%

Fewer clinically relevant adverse reactions when twelve genes are tested before prescribing.

Swen et al., The Lancet, 2023 — PREPARE

The Last Mile — what the problem actually looks like at the bedside. 2:40 · MUSIC: AMBRE — NILS FRAHM

Everything, in four places

Pick what you came for.

Nobody reads a long page. So none of this is a long page — choose the part that matters to you.

Because “too hard” is not a good enough reason.

AI is about to be pointed at a great many problems. This is one of the ones that matters. The knowledge that would prevent a large amount of avoidable harm already exists — it is simply not reaching the person holding the pen.

The reason

The suffering is real and it is measurable. Two in five of the adverse drug reactions that put people in hospital were judged avoidable.Osanlou et al., BMJ Open, 2022
Families wait years for an answer. A rare disease diagnosis takes 5.6 years on average.NHS Genomics Education Programme
Nobody involved is failing. No clinician can read forty million papers. The system asks them to behave as if they had.
This is what the technology is genuinely good at. Synthesis, retrieval and explanation — bounded, cited, and handed to a person to decide.
Compassion is not the opposite of rigour. It is the reason for it. Every refusal on this page exists because someone could be harmed.
This is what the technology should be for. Not another way to sell something, and not a demo. A person makes a better decision because of it.

Why now

The tests are already happening. National pharmacogenomic testing is rolling out now. Results are landing this year, not in five.
Great Britain is roughly a year faster to market than the EU. GB still runs the pre-2017 device rules, so software of this kind is Class I here and Class IIa across the Channel. Same product, shorter route.
The literature grows. The appointment does not. Every year the distance widens between what is known and what can be read inside a consultation.
Nobody has built this part yet. Alerts cover the settled cases. Nothing covers the rest.
We start in the UK. We do not stop there. Every country that sequences hits the same wall — the evidence is the same everywhere, only the rules change.

Built to sit beside the rollout.

National pharmacogenomic testing is already being rolled out, and prescribing alerts are already being built. We are not trying to replace any of that. We are the layer underneath it.

How it fits

No new test. It reads the result you already have.
No new login. The intention is that it opens from where the clinician already is, not beside it.
No new pathway. Nothing upstream has to change.
An alert says what. We show why, and how strong the evidence is.
When there is no rule, there is still evidence. That is our half of the problem.
Nothing is duplicated. Where guidance already exists, we point at it rather than restate it.

Standards and regulation

Today it is not a medical device. It runs on synthetic records and has not been placed on the market or put into service.
Class I when it is. Great Britain runs the UK MDR 2002 rules, not EU MDR. Standalone software of this kind is Class I, self-certified — it drives no other device and does not diagnose.
First evidence comes from retrospective, de-identified data. Clinician-reviewed, so it can start without a prospective trial.
Aligned to the MHRA Software and AI as a Medical Device programme and to FDA Clinical Decision Support guidance.
DCB0129 is the manufacturer’s duty, and it will be ours. The hazard log opens with the first pilot scope — there is no honest way to write one before the clinical pathway is chosen.
DCB0160 sits with the deploying trust. We support our partner through theirs.
DTAC v2 — the refreshed criteria, in force since 6 April 2026.

The record

A hash-chained audit log, verified under test. Tampering is detectable, not merely discouraged.
Every line carries its provenance. Each statement links to the guideline, trial or paper it came from.
The clinician decides, and the decision is recorded. Including when they disagree. An override is data, not an error.
What was suppressed is logged too. Not only what the system said — what it deliberately did not say, and on what grounds.
Role-based access, per-tenant encrypted storage. Built for an organisation, not a laptop.

Three things in. One answer out.

Everything Generesis says is retrieved from published literature and guidelines. Nothing is generated from a model’s memory, and nothing unsourced is allowed through.

What it actually does

It takes the clinician’s diagnosis — the question actually being asked, about this patient, today.
It takes the patient’s pharmacogenomic data from the report that already exists.
It takes only the evidence that bears on both — and leaves out the evidence that is too weak to act on.
It synthesises those three into one ranked, sourced answer that a clinician can read, check and disagree with.

How the evidence is handled

Retracted and predatory sources are excluded by rule. Not filtered afterwards. Not allowed in.
Graded, not scraped. CPIC and PharmGKB evidence levels, trials and case reports carry different weight, and the level travels with the claim.
The quality gate fails closed by design. A claim that cannot be traced to a source does not reach the clinician. There is no fallback to plausible.
Relevance is computed against the actual prescription, not against the genome in general.
Built to work across forty million papers and guidelines. Retrieval-augmented end to end: the corpus is the source of truth, not the weights.

The clinical layer

It reads the pharmacogenomic report today. Full VCF intake against ClinVar, dbSNP, gnomAD, PharmGKB, CPIC and PharmCAT, with ACMG/AMP logic, is the architecture being built onto it.
FHIR bundle import and export work today. SMART-on-FHIR, the EHR note and the prescribing alert are the integration path we are building toward, not yet the shipped product.
It never dumps the genome. Only the gene–drug pairs that bear on this prescription, for this patient, today.
Readable at a glance, defensible on inspection. The summary is one screen. The evidence behind it is one click.

What it refuses to do

It does not diagnose, prescribe, or advise.
It refuses to raise a common variant whose prescribing evidence is weak. MTHFR is suppressed by design, and the suppression is explained.
It says “there is no evidence” out loud. Rather than producing something that reads like evidence.
It never shows a recommendation it cannot source.
It does not act. Nothing happens without a clinician. That is a safety property, not a limitation.
No autonomous diagnosis. No unvalidated outcome claims. Payer return is modelled as a hypothesis, not asserted.

A validation site, and a clinician.

The first study needs no prospective trial and no identifiable patient. That is deliberate: it means a partner can say yes without a two-year approval.

What we are looking for

NHS or provider pilot sites for a retrospective, clinician-reviewed validation on de-identified data.
A clinician who owns the problem. Not a sponsor. Someone who will tell us when the output is wrong, and keep telling us.
No pathway is pre-committed. The first clinical question comes from the clinician, not from us.
One site, one pathway. We would rather be genuinely useful in one place than plausible in ten.
Investors who read the standards first. The regulated route is the moat, not the obstacle.

Munib Rahman, founder

Built Generesis, and activated SEEVAD in 2026 to do it. The architecture is his: a multi-tenant backend, role-based access, per-tenant encrypted storage, and a hash-chained audit log verified under test.
Twenty years getting demanding technical products adopted by organisations that set a high bar for proof before they will let anything near the people they serve — IBM, Goldman Sachs, HSBC, UBS, Samsung, Sky and Kingfisher Group.
Works at the frontier of these tools daily, and has the output to prove it. SEEVAD’s media division delivered a complete AI-produced music video inside a live political campaign’s deadline — work that would once have taken a crew, a budget and six weeks. That fluency is what is now pointed at genomic medicine.
Has run an operation, not only sold into one. At Skyex, a new insurance unit went from nothing to £240,000 of annual revenue inside five weeks.
Not a clinician — but the work is chosen for the reason people become clinicians. To take some avoidable suffering out of the world while the tools to do it exist.
Ran one of the largest microservices practitioner communities in the world — over 2,000 members from international engineering teams.
London.

The most advanced solutions on Earth.

Advanced does not mean clever. It means you can check it — every number on this page has a source, and every decision the system makes can be traced back to one.

Register your interest

Can we help?

Tell us what you need, or what is not working.