Discovering solutions no one thought to look for.
We work on the step almost nobody searches: how a problem is framed before the optimizer ever runs. How it is framed decides the outcome more than how well it is searched. We are building the system that searches it, on the hardest problems scientists and engineers bring us.
Most of the space has never been looked at.
In fields where a single evaluation is expensive (a nonlinear physics solve, a training run, a week of cluster time, a wet-lab experiment), practitioners explore a small fraction of what is available. This is not a failure of rigour. When one trial costs a day, choosing a handful of familiar configurations and stopping at good enough is the correct strategy. The consequence is that most of the space goes unexamined, and the parts nobody examines are exactly where an unfamiliar answer would be.
Three things make it worse. The literature yields hypotheses rather than answers, because published results use different models, baselines and evaluations, so they neither compose nor transfer cleanly. The most consequential choices are made first, under maximum uncertainty, and the feedback often arrives too late to act on. And the judgment that guides those choices is tacit: it takes years to acquire, is seldom recorded, and leaves when people do.
The choices an optimizer cannot make.
What changed is that a machine can now read a literature, propose a formulation and say why it proposed that one. What has not changed is that classical optimizers are far better at numerical search than anything else available, a language model included.
So we separate the two jobs. Semantics, which problem to solve and which candidates are worth considering, belongs to agents. Search, finding the best point inside a stated problem, stays with the Bayesian, evolutionary, multi-fidelity and gradient methods that have decades of work behind them.
What we are actually trying to find out.
These are unsolved. A lab that only publishes its answers is hiding most of its work.
Formulation as a search space
The objective, the fidelity, the parameterization, what to hold constant: decisions like these are treated as context rather than as variables. If they can be named, typed and enumerated, they become a space that can be searched deliberately rather than navigated by habit.
How much of the outcome lives in the formulation rather than in the search?
Deciding without running the experiment
A space of formulations is far too large to evaluate exhaustively. Most candidates must be discarded on evidence that costs a fraction of a full trial. The difficulty is that a cheap test is only informative if it is cheap along an axis the claim does not depend on.
Which choices can be ruled out without paying for the experiment, and how would we know a screen had misled us?
Composition, not method selection
The literature compares methods as alternatives, and asks which one wins. The stronger result is often a sequence: one approach used to reach a starting point that a second could not have found, at a resolution a third made affordable. Under that view the object worth searching is the path rather than the method.
When is a sequence of methods worth more than the best single one, and can that be predicted before running it?
Decision memory
What was assumed, what was ruled out and on what basis, what was overridden and why. This is the layer nobody captures, because it can only be recorded at the moment the decision is made. Experts do not reconstruct their reasoning afterwards, and what they write down later is a summary of the conclusion.
What must be captured at the moment of the decision, because it genuinely cannot be recovered later?
Spending a budget well
Every real search runs under a hard limit: so many solves, so many GPU hours, so many experiments. Stopping at the limit is easy to enforce. Knowing whether the budget was spent on the right questions, when the answer is unknown by construction, is not.
What does it mean to have searched well under a fixed budget, judged without knowing the answer?
Claims with margins, failures with evidence.
This is a field where a claim is easy to make and hard to check. We would rather be provably wrong on the record than plausibly right in a summary.
A prediction before the experiment
A hypothesis states a quantity, a direction, a baseline and a margin specific enough to be wrong, along with the observation that would kill it. Written before the run, not after it.
Failure is the evidence
A method that fails tells you where the boundary is, which is worth more than another success inside it. A failure recorded without diagnostics has already decayed into a missing value.
General by construction, proven one domain at a time
The machinery is built so that it cannot quietly specialise to the first problem it meets. Whether it generalises is then something to demonstrate, not to assert.
Reasoning is part of the result
A result you cannot interrogate is a result you cannot build on.
The first work held to that standard is on the record. We set out to push the published state of the art in solar cell metallization further, and found instead that the designs it holds up as best cannot be printed. Read the report, or see what else we are looking at.
Bring us a problem where the answer is not known.
The research needs real problems, and the hardest ones are not in benchmark suites. We are looking for problems where evaluation is expensive and the formulation is a judgment call, that you know well enough to suspect the standard framing is leaving something on the table, and where you would recognise a genuinely new answer if you saw one.
You bring
- A problem worth solving, and a reason to believe the usual framing is not the only one
- An evaluator we can call, or the means to build one together
- The judgment currently applied by hand, so we can find out how much of it is real
We bring
- A search over formulations, not only over parameters
- Cheap screening, so most of the space is discarded before it costs anything
- A complete record of what was tried, what was ruled out, and why
You keep
- The result, and everything needed to reproduce it
- Your data, which stays yours and stays where it lives
We are looking for people who want to know why.
If you find yourself more interested in why an optimizer returned the answer it returned than in the answer itself, we should talk. Internships and collaborations are remote.
We look less for a background than for a position. You are already solving an optimization problem in a science or engineering domain, you know both its depth and its breadth, and you understand at a high level how agentic systems work. The best project is usually the one you are already carrying, taken apart at the level of how it was framed rather than how it is solved.