Skip to content

Research Engineer, Platform

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 compiler, the dispatch decision and the runtime as one surface, from the moment work reaches the platform to the object that runs on the chip.

Held as production software rather than as one component of any of it: the code generator that takes a block of work down to machine code, the compiler services behind it, the decision about which class of processor each block is offered to, and the runtime that executes it. One person holds all four because a decision about where a computation runs is only worth making if the same hand can emit the object that runs there. The founder wrote nearly all of the platform and reviews the pull requests; no one stands between the engineer and the work, and there is no one to escalate to. This is one of the first engineering hires, so the first responsibility the seat owns is deciding what gets fixed and in what order, and recording why.

Responsibilities

  • The compiler's account of what it can lower is to be made exactly as strong as what it can in fact emit, including the control-flow constructs it sets aside today. The end state is a native path trusted to carry the default route rather than the exception, and that trust rests entirely on the claim: once a block is routed natively there is no second route, so a claim the code generator cannot deliver fails the job rather than degrading quietly.
  • The operations in the public surface that kill the compiler process mid-compile are to be identified, and a crashed compiler is to surface as a named error rather than a bare transport failure. The standard being built toward is that every failure in the compile path names its own cause; compilation runs out of process and the evidence does not outlive the job, so the first step is making the crash visible at all.
  • The native path is to be given a middle end of its own, and the worth of an optimiser settled with a control arm, so that the speed of a natively generated block comes from a pipeline this company owns rather than from a stock optimisation level applied further down the stack. There is little of that pipeline today beyond what preserves semantics, and the code generator targets a generic processor rather than the machine the work lands on.
  • The runtime's memory model is owned in this seat, so that the size of a job a customer can submit is a property of the problem rather than of the process that happens to be serving it. A single large quantum-chemistry job can exhaust the memory of a large-memory machine and release none of it while the process lives; buffer planning and the scheduling of frees are the levers, and both belong here.
  • The integrated cross-hardware demonstration is to be delivered, after which chip classes continue to move from modelled to dispatchable until the coverage reaches the silicon that matters. Only a small minority of the classes this company models are dispatchable today, each of the others is blocked for a reason already written down, and the bar is a real run whose record shows the hardware was used, where 'the run completed' is not evidence.

You may be a good fit if you

  • The candidate will have shipped a compiler, runtime or accelerator toolchain that other people depended on: MLIR, LLVM, XLA, IREE, Triton, a vendor stack, or an in-house system no one has heard of that worked.
  • The candidate reads Rust at depth, and can open a large unfamiliar module, add a case to it, and identify which stated contract that case now rests on.
  • Fluency in LLVM at the IR-text and TargetMachine level is required: a pass-builder pipeline is distinguished from a code-generation optimisation level, and an undefined symbol at dynamic load is familiar ground rather than a novelty.
  • Competence in modern C++ is required inside a substantial long-lived service, not only in a short benchmark.
  • The successful candidate distinguishes a median from a measurement: the workload and the size are stated beside every number, and what a result does not establish is recorded.
  • The candidate will have worked best unmanaged, and will want a surface rather than a ticket queue.
  • The engineer works with coding agents as a matter of course and verifies what they produce; that is how this codebase was built and how it will continue to be built.

Strong candidates may also

  • Command of MLIR internals as working vocabulary: bufferization, unrealized_conversion_cast, scf.for tensor carries, destination-passing style, Linalg identity maps.
  • Command of the alternatives to MLIR as well: the separation of compute from schedule with cost-model-driven autotuning, of the TVM class, or the tile-and-block model with explicit warps, stages and shared-memory staging, of the Triton and Pallas class.
  • GPU work at the symbol level: device-code targets and their assembly at link time, solver and BLAS symbol conventions, and the class of defect in which one sticky device error poisons a long-lived process.
  • Reverse-mode automatic differentiation over a computation graph with structured control flow, which is the kind of capability that opens a new field of use rather than repairing an existing one.

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.