Researcher, Post Training

About the role

As a Researcher in Post Training, you will shape how nyra labs models behave after pretraining.

You will develop methods that make speech models more accurate, controllable, robust, and aligned with what people actually need. This includes post-training recipes, feedback-driven learning, data curation, model evaluation, and the systems required to run reliable experiments at scale.

This is a research role with strong engineering ownership. You will take ideas from an initial hypothesis through experimentation, evaluation, and release.

Why we need you

Pretraining creates capability. Post training determines whether that capability becomes useful.

Speech models need to understand what should be preserved, how uncertainty should be handled, and how behavior should change across tasks, languages, speakers, and clinical contexts. Generic alignment methods rarely account for the details that matter in real speech: hesitations, repetitions, interruptions, atypical pronunciation, silence, and incomplete utterances.

nyra health gives us access to a uniquely large, therapist-labeled dataset of neurological speech, including millions of recordings from real clinical settings. You will help turn that asset into models that behave reliably for people underserved by existing speech technology.

About the company

At nyra health, we build software that supports clinics, therapists, and patients throughout neurorehabilitation. myReha delivers personalized therapy, while nyra insights helps clinical teams manage and understand patient progress.

nyra labs is the research arm of nyra health. We turn difficult problems encountered in practice into open models, datasets, benchmarks, and research that the wider community can build on.

If that resonates with you, we would love to hear from you.

What you’ll shape

• Post-training methods: Develop and test approaches including supervised fine-tuning, preference optimization, distillation, feedback-driven learning, and reinforcement learning where useful.
• Training data: Create high-quality post-training datasets through curation, annotation, synthetic data generation, and model-assisted data improvement.
• Model behavior: Improve instruction following, verbatim transcription, uncertainty handling, long-form consistency, multilingual performance, and resistance to hallucinations.
• Evaluation: Build benchmarks and failure taxonomies that reveal whether models are genuinely improving.
• Experimental systems: Implement, debug, and scale training and evaluation pipelines with a strong focus on reproducibility.
• Specialized models: Work with research and product teams to adapt foundation models to specific speech and clinical use cases.
• Open releases: Contribute to publications, model releases, datasets, and technical reports.