Job Description
We're hiring 2 engineers for a Jamba-JEPA hybrid (toward JEPA-Reasoner). Prioritize research-to-code ability, not general 'ML engineer' vibes.
Role A (Research Engineer – JEPA/representation): implemented self-supervised / latent objectives before (EMA targets, collapse avoidance like BYOL/VICReg/Barlow), can design ablations for leakage/bypass/trivial solutions.
Role B (Lead ML Engineer – backbone integration): strong PyTorch + Hugging Face, has modified pretrained LLM forward passes (prefix embeddings/adapters/conditioning), can integrate SSM/Transformer hybrids.
Non-negotiable screen: ask for 1 repo where they implemented a paper idea (new loss/arch) + ran ablations (not just finetuning).
5 questions to ask:
Link a repo where you implemented a paper idea end-to-end.
Have you modified a pretrained HF model's forward pass? Example.
How do you prevent collapse/leakage/bypass in latent prediction?
...
Apply for this Job
Submit your application for the AI Software Engineer position at MHZ TimeCloak.
Apply Now Save for Later