Skip to content

Research Engineer, Model Training & Evaluation

Zurich, CH · hybrid

About ORIQX

Compute is fragmenting. For a decade the answer was one CPU and one class of GPU. It is now many: further GPU classes, gate quantum processors, annealers, analog and thermodynamic samplers, wafer-scale parts and custom accelerators, each an order of magnitude better or worse depending on the problem, and each with a different cost, energy and error profile. The decision of what to run where is currently made once, by hand, and is wrong as soon as prices, hardware or problem size change.

ORIQX is three things: a model, an engine and hardware. The ORIQX compute model is being built, starting now: a model that works in the vocabulary of a domain and states what to compute, never where. The ORIQX compute engine is live today: before anything executes, it predicts what a computation will cost in time, money, accuracy and energy on every class of chip we model, runs each part where it wins, and returns a report of predicted against actual. ORIQX hardware is being simulated now, and its blueprint follows in twelve months: the design of a heterogeneous rack built from other vendors' silicon, sized to a real portfolio of workloads.

The engine sits beneath the orchestrators rather than among them, and all three are meant for every vendor's silicon rather than one.

About the role

This seat owns the ORIQX compute model's training runs and the evaluation that judges them: the held-out domain problems, the comparison against general models writing code, and the regression suite that keeps a better model from being a worse one on last month's problems.

The ORIQX compute model is being built, starting now, and a model is only as good as the measurement that says so. This seat owns that measurement and the runs it measures: the training runs themselves, on whichever class of silicon suits each one; the evaluation harness of held-out domain problems, scored by executing the output; the comparison against a general model writing code on the same hardware budget; and the regression suite that stops a new checkpoint shipping when it is better on average and worse on problems the last one solved. It works beside the model's first seat, which decides what the model states and what it is trained on. This seat decides whether a run was an improvement, and it is the one that says no. The founder reviews the work directly.

Responsibilities

  • Make every training run reproducible from a clean checkout, with the data, the configuration and the hardware recorded beside the result, so that any number the model is judged by can be regenerated by somebody who was not there.
  • Build the evaluation harness: held-out domain problems, output scored by executing it rather than reading it, variance across seeds reported beside every mean, and a held-out set checked for leakage from the training corpus before it is trusted.
  • Run the comparison against general models writing code on every model version, with the same hardware budget and the same number of attempts on both sides, and report the result internally whichever way it goes.
  • Stand up the regression suite: a new checkpoint that improves the average and fails problems the previous one solved does not ship until somebody has decided, in writing, that the trade is worth it.

You may be a good fit if you

  • The successful candidate has owned training runs for a model other people depended on, and can describe a run that looked like an improvement and was not.
  • The engineer has built an evaluation harness or benchmark suite that caught a regression before it shipped, and can say what it measured and what it missed.
  • Python and PyTorch are daily tools, distributed training on more than one class of accelerator is familiar ground, and the training configuration is treated as code that gets reviewed.
  • The candidate reports variance, attempts and baselines beside every number, and is uneasy about a single-run result.
  • Coding agents are used as a matter of course and what they produce is verified; that is how this platform was built and how it will continue to be built.

Strong candidates may also

  • Experience evaluating code-generation models by execution, on benchmarks such as HumanEval or KernelBench, and knowledge of where those benchmarks leak.
  • Familiarity with training-data contamination: how to detect it, and what it does to a published result.
  • A computational-science background, so that a domain problem's correctness can be judged without a specialist in the room.

Logistics

Education
A degree in computer science, electrical engineering, physics or mathematics is one way to be ready for this work. Equivalent experience is another, and the two are weighted equally. The assessment is what the candidate has built.
Location
Based in Zurich, or hybrid. Relocation support.

Apply even if you do not meet every line above. Few strong candidates meet all of them, and the ones who exclude themselves early are disproportionately the people we would want to have met.

Come work with us

What you can do
We care about what you can do, not where you learned to do it.
Compensation
Every seat has a real band, and it is given on the first call: competitive for a company at this stage, and not matched to a frontier laboratory's cash.
Equity
Equity is part of every offer. The instrument, the vesting and the Swiss tax treatment are settled in the offer letter, with the candidate. No range is published, because none is agreed for a seat until then.
Where
Based in Zurich, or hybrid. Relocation support.