Four problems on our desk.
We are hiring a Controls Lead. Below are four problems we have not solved. None is a puzzle with an answer we are holding back. If you want to argue with one of them, that is the application.
What we would like an argument about.
-
Is the load actually periodic, and does the periodicity survive the power supply?
A training cluster's rhythm lives in the collective library and the training framework. It does not live in the job scheduler. That rhythm is recoverable from measured power at the point of common coupling, by autocorrelation and phase-lock.
What nobody here knows is how much of the structure survives the PSUs and the DC-link capacitance on the way to the AC side. How would you measure it? What would you accept as proof either way?
-
What does a controller-in-the-loop bench have to reproduce to be worth its cost?
We are specifying the bench now. It has a purchase order against it and a date. The question is what it has to reproduce faithfully before it tells us anything we did not already assume.
What would you refuse to buy it without?
-
Where does a rig stop finding defects in the code and start finding them in the model?
Our answer so far is two lanes and one rule. The regression lane is seed-locked with fixed gates. The exploration lane is randomised and checks invariants only. Every assertion carries a sensitivity proof: a deliberate mutation that turns that assertion red. An assertion that cannot be made to fail is not testing anything.
Where does that method still leave a blind spot?
-
What is the honest way to separate a controller's contribution from the hardware it commands?
Ours is a frozen-controller counterfactual: the same hardware, the same scenarios, the controller held still. It is scheduled, and we think it is the weakest of the four answers on this page.
Tell us a better one.
What the Controls Lead owns.
-
The controller-in-the-loop bench
Specification through commissioning. Purchase order November 2026. You decide what it has to reproduce and you sign it off.
-
The turbine governor and AVR model
Negotiated into the framework supply agreement with the OEM packager from September 2026.
-
The non-stiff grid fixture
A parallel lane in the regression network, running from the next release.
-
The Stage 0 measurement rig
Problem 01 is yours to close, and this is the instrument. You define what would settle it, you build the rig, and you publish the answer either way. It happens before the bench is bought, because a bench fed an assumed load profile validates the assumption and nothing else.
Stated plainly.
Munich. On site. The bench is a physical object.
Meaningful, with vesting and a cliff. Discussed directly, not published.
Pre-seed. The round is €2M and open.
How to reach us.
Write to Matthias@batteryvault.de with an argument about one of the four problems above. A CV is welcome, and it is not what we will read first.
Speculative applications
We have one open role. If you work in power systems or controls and one of the four problems is yours, write anyway, the same way. We would rather know you now.