We are early, and we say so. That makes the useful shape of a partnership specific: a problem where the volume of experiments is the bottleneck, a partner who can verify our work independently, and a result both sides can publish.
Different commitments, different timelines, same requirement: something measurable at the end.
We bring the loop, the harness and the compute; you bring the domain problem, the objective and the ability to say whether the answer is meaningful. Methods are published; domains are not, unless the parties agree.
Best first step: a scoping note describing the objective and what a good result would look like.
For organisations with a computational bottleneck in a domain we have not worked in — materials screening, treatment design, instrument calibration. This is deliberately the later stage of our roadmap, and we will say so if the fit is premature.
Best first step: a description of the loop you run today and where its turns are lost.
An automated lab consumes compute in a pattern that is unlike a training cluster: thousands of small, short, interruptible jobs with strict per-trial ceilings. If you run capacity that suits that pattern, we would like to talk.
Best first step: the shape of what you offer — regions, interruption model, minimum useful job size.
For researchers who want to work on automation without leaving their institution. Joint supervision, published outputs, and honest scoping about what can and cannot be shipped.
Best first step: see careers and mention the arrangement you have in mind.
Both columns are real constraints. A partnership works when the second one is satisfiable, not when the first one sounds impressive.
Four steps, each small enough to stop at. Most conversations end after step two, and that is a good outcome — not a rejection.
Forty-five minutes. You describe the loop you run today and where its turns are lost; we say honestly whether automation is the relevant fix.
One page, written jointly: objective, what a good result looks like, what would make us stop. If this page cannot be written, the project is not ready.
Harness, baselines, held-out objective, budgets and review gates. The unglamorous step that decides whether anything afterwards is worth believing.
Bounded in time and spend, with a pre-agreed stopping condition. Success is a result you can verify — or a clear, publishable account of why the approach did not work here.
Steps one and two cost nothing and carry no obligation. We are not going to ask for a signature before we have said something useful.
Five questions. If the answer to any of the first three is no, automation of this kind will not help yet — and we would rather say that early than discover it expensively.
| Question | Yes looks like | No looks like |
|---|---|---|
| Is there a measurable objective? | A number or a decision that two independent labs would agree on | “Better”, “more promising”, or a judgement that only an expert can make in context |
| Are trials parallelisable? | Hundreds of independent variations that do not depend on each other's results | Each experiment determines the next one, so there is nothing to run in parallel |
| Is the computational stage the bottleneck? | Design, screening or analysis consumes the calendar time | Waiting on synthesis, instruments, approvals or clinical follow-up |
| Can the result be verified independently? | You can reproduce or refute our claim without our involvement | The result is only meaningful inside our system, on our data, by our metric |
| Is publication acceptable? | The method at minimum can be described publicly, with a dated account either way | The existence of the collaboration itself cannot be mentioned |
Two “no” answers out of the first three usually means the right advice is to spend the next year on measurement rather than automation.
Stated plainly, because these are the requests we get most often and the answer will not change on the third email.
A short description of the cycle you run today is more useful to us than a polished proposal. We will tell you honestly whether we can help.