Constellation is building a foundation model of human state, translating multimodal human data into generalizable embeddings to power clinical, robotic and AI systems. It is a venture-backed team of 15 to 20 researchers and engineers based in San Francisco's Mission District.

About the role

Posted as Founding Applied Scientist, this is a hybrid research and engineering role working with what the company describes as the richest multimodal dataset ever collected. You will architect pipelines transforming that data into useful features, define the ground truth, design the benchmarks measuring progress, and deploy the models that power the product. The role also carries full visibility into the data lifecycle and influence over collection strategy.

What you will do

• Train and fine-tune deep learning models in PyTorch or JAX, translating raw signals into interpretable behavioral constructs.
• Perform advanced time-series analysis identifying latent state changes, clusters and transitions.
• Architect cloud-based pipelines capable of processing petabytes of data.
• Define ground truth and evaluation metrics, curating validation data and rigorous benchmarks.
• Implement active learning loops improving model performance against massive incoming data streams.
• Clean, wrangle and interpret noisy, high-variability real-world data.

What they are looking for

• Top-tier publications at NeurIPS, ICLR, CVPR or ICML, or in high-impact journals.
• Proficiency in PyTorch or JAX, with the ability to implement papers from scratch and modify architectures.
• Strong Python expertise and deep experience in time-series analysis, plus at least two of computer vision, NLP/audio/speech, or biosignal and neural time series.
• Experience architecting data processing pipelines and deploying models in the cloud.
• Nice to have: neuroadaptive systems, biofeedback or closed-loop stimulation paradigms, and exposure to large-scale foundation models.