Research Engineer

Constellation•San Francisco, CA
•Onsite

About The Position

Constellation is creating an AI-human translation layer and building foundation models from a rich multimodal dataset. The Research Engineer will work between data and models to accelerate the research and training loop. This role involves orchestrating and optimizing training runs on long-horizon multimodal sequences, building pipelines that transform a growing corpus into model-ready data, and developing reliable research code, including libraries, dataloaders, and evaluation harnesses. The engineer will collaborate closely with research and engineering teams and help turn prototypes into repeatable pipelines while preserving flexibility for exploration.

Requirements

  • At least 3 years building ML systems or research infrastructure, including distributed training runs on 100 or more GPUs.
  • Expert-level Python skills and deep knowledge of PyTorch internals, including DDP and FSDP, mixed precision, gradient accumulation, and profiling.
  • A track record of building or maintaining research libraries used by others.
  • Experience with data pipelines for large unstructured and multimodal datasets.
  • Familiarity with columnar and streaming formats such as Zarr, Parquet, Arrow, Lance, WebDataset, or Vortex, and their tradeoffs for random access and sequential throughput.
  • Hands-on experience with experiment-tracking and dataset-versioning tools such as ClearML, Weights & Biases, or MLflow.
  • Research fluency sufficient to read a paper, reimplement a component, and assess whether a loss curve is broken.
  • Ability to debug issues ranging from corrupted shards and incorrect collate functions to training runs that diverge.
  • Care for the well-being of people affected by the technology and ability to thrive in a collaborative, high-bandwidth environment.

Nice To Haves

  • Custom kernel work in CUDA or Triton, or compiler-level optimization with tools such as torch.compile, TensorRT, or ONNX.
  • Experience with time-series or multimodal data requiring synchronization across streams.
  • Experience with low-latency or edge inference, quantization, or distillation.
  • Comfort using Rust or C++ when Python is not the right tool.
  • Interest in neurotechnology, mental health, or human-AI interaction.

Responsibilities

  • Orchestrate and optimize distributed training runs, including configuration, mixed precision, checkpointing, recovery, and profiling.
  • Build high-throughput data-loading systems for terabyte-scale multimodal data using sharding, prefetching, caching, and appropriate data formats.
  • Align and synchronize multi-stream time-series data, adapt to changing collection protocols, and manage dataset versioning and lineage.
  • Build reproducible feature-generation pipelines for raw recordings, ensuring derived datasets are versioned and can be recomputed when upstream data changes.
  • Develop clean, well-tested, well-documented research libraries for datasets, models, transforms, and metrics.
  • Ensure each training run is traceable through its code version, configuration, dataset version, and environment.
  • Build evaluation harnesses that run automatically on new checkpoints and identify regressions.
  • Turn research prototypes into repeatable pipelines without limiting researchers’ ability to explore.

Benefits

  • Comprehensive health, dental, and vision insurance with premiums fully covered.
  • Renovated, light-filled office in San Francisco’s Mission District.
  • Relocation assistance for employees joining in San Francisco.
  • Workspace setup stipend.
  • Monthly team dinners and outings, plus twice-yearly off-sites and retreats.
  • Fully stocked kitchen with snacks.
  • Travel sponsorship for relevant conferences and regular professional development activities.
  • At least 12 weeks of fully paid parental leave with a soft-landing transition back.
  • Paid time off (PTO).
© 2026 Teal Labs, Inc
Privacy PolicyTerms of Service