Find the question
that matters.
Start from a concrete failure or constraint. Make the experiment clear enough that its result changes what you build next.
Careers / Build what comes nextThink deeply.
Think deeply.
Build all the way.
Help turn a different idea of model memory into systems people can use. There is work to do from the first equation to the last kernel.
01 / The kind of workAsk precisely Build completelyCross the
Share the substanceMake the work
A useful idea should survive contact with a benchmark, a debugger and a real device.
Find the question
Cross the
abstraction layers.
Move between model behavior, training, runtime and hardware. Follow the bottleneck to the place where it can actually be solved.
Make the work
inspectable.
Write the method, preserve the conditions and publish useful artifacts. Good engineering leaves the next person a clear starting point.
02 / Future team directionsBring your
Bring your
particular obsession.
Model memory, training dynamics, efficient execution, or the machine beneath it all. Explore where your work could connect.
Research Scientist — World-State Models
Research / Memory + reasoningInterest welcome
Research Scientist — World-State Models
Research / Memory + reasoningStudy what a compact model should remember, how it should write memory, and when retrieval helps it reason. Turn architectural ideas into controlled experiments and useful open models.
Questions you could work on
- Learned memory admission, replacement and retrieval under a fixed state budget.
- Long-running reasoning and multimodal state representations.
- Experiments that distinguish useful recall from accidental benchmark shortcuts.
Work that would help us understand you
A paper, an experiment, a model implementation or a careful negative result. We value clear questions and strong empirical judgment.
Share your researchResearch Engineer — Training + Evaluation
Engineering / Models + experimentsInterest welcome
Research Engineer — Training + Evaluation
Engineering / Models + experimentsMake research repeatable at the scale where model behavior becomes visible. Build the training and evaluation work that connects a promising mechanism to a model people can inspect and use.
Work you could help shape
- Training pipelines for memory writing, admission and retrieval.
- Evaluation across reasoning, code, vision and long-running workloads.
- Reliable checkpoints, clear experiment records and reproducible releases.
Work that would help us understand you
Training code, an evaluation harness, a model adaptation, or a difficult experiment you made reproducible. Tell us where it failed and what you learned.
Share what you builtInference Systems Engineer
Systems / CUDA + deploymentInterest welcome
Inference Systems Engineer
Systems / CUDA + deploymentMake the model’s execution match its architectural promise. Work on the kernels, memory layouts and serving paths that turn fixed history into practical throughput and latency.
Work you could help shape
- RMA and recurrent-state kernels, graph execution and GPU memory behavior.
- Precision-aware optimization with correctness and allocation checks.
- Runtime integration, quantization and constrained-device deployment.
Work that would help us understand you
A kernel, a profile, a runtime contribution or an optimization with a measured before-and-after result. Good comparisons matter as much as a fast number.
Share a systems projectHardware Architect — Model-to-Silicon
Architecture / Specialized computeInterest welcome
Hardware Architect — Model-to-Silicon
Architecture / Specialized computeExplore what changes when a model and its execution substrate are designed together. Connect fixed-weight, fixed-state computation with realistic architecture and implementation constraints.
Questions you could work on
- Dataflow, memory organization and token-level execution schedules.
- Cycle models connected to area, power and physical implementation.
- Numerical behavior and end-to-end verification of specialized compute paths.
Work that would help us understand you
An architecture study, RTL, a simulator or a hardware-software project. Show how the model and the machine informed each other.
Share your perspectiveThese are prospective team directions. A conversation does not imply a confirmed vacancy or an offer. Use a short introduction and links to work you can share publicly; engagement details are discussed directly.
Show us what you built.
Tell us what still breaks.
A project, a paper or an unusually good question is a useful place to start.