Oxiedo is pre-revenue and unfunded, and one person has built everything that exists. The commitment shape is stated on every role below — some are equity, one is a single consultation, one is a research partnership with no money attached in either direction — because it would be easy to make this list look larger than it is, and you would find out later.
They are ordered by how much each one unblocks, not by seniority. The first is not a job at all, and it is the one that matters most.
- Stage
- Pre-revenue, pre-customer
- Team
- One
- Location
- Remote, with the headquarters relocating to San Francisco
- Reply time
- A few days, usually less. Every message read by a person
- 01
Dataset partner
The binding constraint on everything else
No clinical, biological, financial or defence data has ever touched this system.
Every result to date is CIFAR-10 or CIFAR-100. That is not a gap in the engineering — the engineering is done and published. It is a gap in access, and it is the single thing most limiting this company.
If you hold a labelled dataset you suspect is dirty, or you have had a training run fail in a way nobody could explain afterwards, that is the conversation we want.
A dataset is not a gift and we do not ask for one. It is a research partnership: we run the diagnostic, you receive the report, your data never leaves your environment unless you decide otherwise, and we publish together if you want to.
Apply for this - 02
Co-founder
The architecture exists. What is missing is a second person.
Someone who has stood on the other side of the buying table in one regulated domain — inside a hospital's AI group, a bank's model risk function, a biotech's data team, or coordinating a multi-site research consortium.
Technically literate enough to survive a CTO's questions without needing to be the architect. What matters more is a network that cannot be built from where this is being built from.
This is not a hire. It is the other half of the company, and it is described that way because pretending otherwise would waste a year of someone's life.
Apply for this - 03
Researcher
Four measurements decide how much of this is real, and not one has been run.
Two of them matter most. Does the per-component health signal actually lead the global loss curve, and by how much? And does inter-component disagreement separate mislabelled data from data that is merely hard?
Four of the seven applications depend on the second answer. A negative result narrows what the product can be, which is exactly why the study is scheduled before the build rather than after it.
Beyond those: the architecture is demonstrated on fully-connected, convolutional and residual families. Whether it generalises to transformers is genuinely open — not assumed and not dismissed.
You would run experiments capable of falsifying the thesis, and co-author the result whichever way it goes.
Apply for this - 04
Regulatory consultant
It gates the largest application in the product, and it has not started.
Someone who has actually filed a change-control plan for a machine learning system — under the FDA's PCCP framework, the EU AI Act, or an equivalent regime. Filed one, not read about one.
The work is to read a specification and say whether the artefact format matches what a reviewer expects to receive, or whether it is something adjacent that a reviewer would not recognise.
Either answer is worth having. If it is the wrong shape, that is worth knowing before more is built on it.
Apply for this - 05
Data-protection lawyer
Until this is answered, certified deletion is a strong story rather than a certifiable product.
Someone working in AI data rights, willing to state what may and may not be put in writing when data is removed from a trained model.
The specific question: if we delete a bounded structure trained on a named sample set, and we can state a measured bound on the residual influence of that data — what may we certify, what may our customer certify to theirs, and does any of it constitute erasure under Article 17?
A certificate whose wording has not been checked is worse than no certificate, because it creates liability instead of removing it.
Apply for this
There is no form, no tracking system and no process.
Write to us and say which one, and what you have done that is relevant. A paragraph is enough. A CV is fine if you have one to hand, and not required if you do not.
If it is not a fit we will say so directly rather than going quiet, because the alternative wastes your time and we would rather not. If it is, the next conversation is with the person who built this, because there is nobody else.
One form, and it goes straight to the founder.
No tracking system, no automated screen, no acknowledgement email written by somebody else. If it is not a fit you will be told directly rather than left to work it out from silence.