Research Engineer, Compute Model
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 starts the ORIQX compute model: a model that works in the vocabulary of a domain and states what to compute, never where, so that the engine can place the work on whichever class of silicon wins. The model is not built, and this seat builds it.
A general model asked to solve a heavy computation writes code in a programming language, and somebody still has to run it, debug it and tune it for each chip, by trial and error. The ORIQX compute model is meant to remove that loop. Given a problem in a domain's own terms, it states what has to be computed and leaves where it runs to the engine, which is live today, prices every part of a workload before it runs and places each part on the silicon where it wins. This seat owns the model end to end: what it states, what it is trained on, how it is trained, and the evaluation that decides whether it is better than a general model writing code. The training corpus starts from the engine's own runs and is small today. A second seat, for the training runs and the evaluation harness, is funded to start shortly after this one. The founder sets direction and reviews the work, and nobody stands between the engineer and the decisions.
Responsibilities
- Decide what the model states, and show that the engine can use it. Its output has to be something the engine can price and place without anybody rewriting it for a chip. The first deliverable is that boundary written down, with a set of domain problems carried across it end to end.
- Build the training corpus from the engine's own runs, and say plainly how small it is. Every job the engine runs records what was predicted and what was measured; turning that record into training data, without letting the evaluation set leak into it, is most of the quiet work of the first year.
- Stand up the comparison the seat exists for: on held-out domain problems, the model against a general model writing code, both run on the same hardware budget and scored by executing the output. The result is reported whichever way it goes.
- Keep every claim about the model at its evidence. Until the evaluation reports, the company says only that the model is being built; this seat owns the first sentence that says more, and the measurement behind it.
You may be a good fit if you
- The successful candidate has trained or fine-tuned a language model on a corpus they assembled themselves, and can say what was wrong with the corpus and how they found out.
- The engineer has built an evaluation other people trusted: held-out problems, a scoring rule that executes the output rather than reading it, and a record of the time the evaluation itself was wrong.
- The candidate has worked on code generation, program synthesis or a model whose structured output a machine then acts on, and knows the difference between output that looks right and output that runs.
- Python and PyTorch are daily tools, Rust is read well enough to follow a problem into native code, and the candidate is at ease building next to an engine somebody else wrote.
- The candidate treats a benchmark result as a claim with a boundary: which problems, which baseline, how many attempts, on what hardware.
- 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
- A background in a computational domain such as chemistry, simulation, optimisation or finance, so that a problem stated in the domain's own terms is legible without translation.
- Familiarity with how general models perform on code benchmarks, KernelBench among them, and with where those benchmarks stop measuring what they claim to.
- Experience with reinforcement learning from execution feedback, or with a model whose reward came from running its output.
- Reading knowledge of compiler infrastructure such as MLIR or LLVM, enough to follow a problem into native code.
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.