Research Engineer

Open role

San Francisco (On-site), Full-time

Constellation is creating the AI-human translation layer that ensures humanity evolves alongside our technology. Our mission is to leverage AI towards addressing deep and meaningful problems at the core of the human experience: empowering people towards their goals, augmenting our cognition and emotional wellness, and understanding ourselves and each other. Our path forward is to move away from AI that captures human knowledge towards AI that truly understands what it is to be human.

We are generating the richest multimodal dataset ever collected to build a new class of foundation models and we're seeking the team of researchers that will build them.

We're looking for a Research Engineer to sit between our data and our models and make the whole loop faster. You will orchestrate and optimize training runs on long-horizon multimodal sequences, build the pipelines that turn a messy, daily-growing corpus into something our models can learn from, and write the research code (libraries, data loaders, evaluation harnesses) that lets the rest of the team try ideas quickly and trust what they see.

You'll work closely with both our engineering and research teams.

• Orchestrate and optimize training: distributed configuration (DDP/FSDP), mixed precision, checkpointing and recovery on multi-day runs, and profiling to find which of kernel, data loader, communication or pipeline is actually the bottleneck before touching anything.

• Build high-throughput data loading over TB-scale multimodal data: sharding, prefetching, caching and format choices that keep GPUs saturated rather than waiting on I/O.

• Wrangle complex data: align and synchronize multi-stream time series, handle changing collection protocols, and own dataset versioning and lineage as the corpus grows.

• Build pipelines that derive rich features from raw recordings so every derived dataset is versioned, reproducible from its inputs, and cheap to recompute when upstream data changes.

• Write research code that makes research better: clean, well-tested, well-documented libraries for datasets, models, transforms and metrics that the team builds on rather than around.

• Make every run traceable: code version, config, dataset version and environment recoverable from any result.

• Build evaluation harnesses that run automatically on new checkpoints and surface regressions before anyone goes looking.

• Turn research prototypes into repeatable pipelines without flattening the flexibility researchers need to keep exploring.

• 3+ years building ML systems or research infrastructure, including distributed training runs on 100+ GPUs.

• Expert-level Python and deep knowledge of PyTorch internals: DDP and FSDP, mixed precision, gradient accumulation, and the profiling tools to tell a slow model from a starved one.

• A track record of building or maintaining research libraries others depend on. Contributions to packages like , torch audio, torcheeg, , neural sets, or comparable internal tooling are exactly what we're looking for.

• Experience with data pipelines over large unstructured and multimodal datasets, and familiarity with columnar and streaming formats (Zarr, Parquet, Arrow, Lance, Web Dataset, Vortex) and the tradeoffs between random access and sequential throughput.

• Hands-on experience with experiment tracking and dataset versioning tooling (ClearML, Weights & Biases, MLflow, or similar).

• Enough research fluency to read a paper, reimplement a component, and tell whether a loss curve is broken.

• Debugging range from a corrupted shard to a silently wrong collate function to a training run that quietly diverged on day two.

• You care about the well-being of the people this…