Machine Learning Systems & Infrastructure Engineer

We’re looking for bold, innovative individuals driven by a passion for tackling hard problems in generative 3D AI. You should thrive in an environment where creativity meets technical challenge, take pride in craft, and collaborate closely with a small team building frontier systems. We are seeking a Machine Learning Systems & Infrastructure Engineer to build and own the systems that turn raw real-world data into trained world models and reliable production endpoints. You will design, implement, and operate scalable training stacks, data ingestion pipelines, experiment orchestration, and model serving for large diffusion-based generative models. The role is hands-on and code-heavy — you will work inside the same monorepo as the research team, mostly in Python, and should be as comfortable refactoring a trainer class or a dataset loader as you are writing Terraform. Responsibilities

- Distributed training enablement: Improve high-throughput training stacks (e.g., PyTorch DDP/FSDP, NCCL) for performance, stability, and reproducibility, including preemption-safe and sharded checkpointing. - Data systems and pipelines: Build end-to-end Python pipelines that turn third-party capture sources into clean, versioned training datasets — including scraping (e.g., Playwright) and preprocessing — and optimize the underlying storage at petabyte scale (object storage, fuse mounts, caching layers, shared filesystems, and relational / analytical / embedded metadata stores). - ML workflow orchestration and serving: Operate the systems researchers use to launch experiments, data jobs, and production endpoints — workflow engines (e.g., Kubeflow Pipelines, Airflow), GPU schedulers (e.g., Volcano, Slurm), experiment trackers (e.g., MLflow, Weights & Biases), and managed-inference platforms (e.g., Modal, Triton) — and maintain a launcher SDK for one-command runs.

Key Qualifications - 3+ years writing production-quality Python in a large, multi-author codebase, with strong SWE fundamentals (ML systems experience strongly preferred). - Hands-on with modern ML training stacks (PyTorch; DDP/FSDP or comparable); have personally debugged distributed jobs across many GPUs and nodes.

- Proficient with containers (Docker, Kubernetes) and comfortable reading and writing IaC (Terraform) for the surfaces you ship. - Strong working knowledge of how to store and query large datasets at scale: SQL fundamentals; relational (e.g., Postgres), analytical (e.g., BigQuery, Snowflake), and embedded (e.g., SQLite) stores; and object storage with caching layers. Familiarity with ML workflow orchestration and experiment tracking (e.g., Kubeflow Pipelines, MLflow). - Experience with monitoring and observability tooling (e.g., Prometheus/Grafana, OpenTelemetry) and CI/CD for infra and ML workflows (e.g., GitHub Actions).

- Hands-on GPU compute and performance debugging (CUDA/NCCL, GPU utilization, networking bottlenecks, profiling). - Working knowledge of cloud environments (AWS, GCP, or Azure), including object storage, IAM, and cost awareness. - Proficient with containers (Docker, Kubernetes) and comfortable reading and writing IaC (Terraform) for the surfaces you ship.

- ML workflow orchestration and serving: Operate the systems researchers use to launch experiments, data jobs, and production endpoints — workflow engines (e.g., Kubeflow Pipelines, Airflow), GPU schedulers (e.g., Volcano, Slurm), experiment trackers (e.g., MLflow, Weights & Biases), and managed-inference platforms (e.g., Modal, Triton) — and maintain a launcher SDK for one-command runs. - Containerization and packaging: Ship workloads with Docker and Kubernetes; maintain IaC (Terraform) for the surfaces you own and CI/CD pipelines, including self-hosted GPU runners. - Observability and reliability: Monitoring, logging, and alerting for job performance, data-pipeline health, and cost (e.g., Prometheus/Grafana, OpenTelemetry); define SLOs and incident response for the systems you own.

- Working knowledge of cloud environments (AWS, GCP, or Azure), including object storage, IAM, and cost awareness. - Proficient with containers (Docker, Kubernetes) and comfortable reading and writing IaC (Terraform) for the surfaces you ship. - Strong working knowledge of how to store and query large datasets at scale: SQL fundamentals; relational (e.g., Postgres), analytical (e.g., BigQuery, Snowflake), and embedded (e.g., SQLite) stores; and object storage with caching layers. Familiarity with ML workflow orchestration and experiment tracking (e.g., Kubeflow Pipelines, MLflow).

- Containerization and packaging: Ship workloads with Docker and Kubernetes; maintain IaC (Terraform) for the surfaces you own and CI/CD pipelines, including self-hosted GPU runners. - Observability and reliability: Monitoring, logging, and alerting for job performance, data-pipeline health, and cost (e.g., Prometheus/Grafana, OpenTelemetry); define SLOs and incident response for the systems you own. - Security and access: Manage secrets, IAM, and network boundaries (e.g., Tailscale, cloud VPC) for the systems you own.

- Strong working knowledge of how to store and query large datasets at scale: SQL fundamentals; relational (e.g., Postgres), analytical (e.g., BigQuery, Snowflake), and embedded (e.g., SQLite) stores; and object storage with caching layers. Familiarity with ML workflow orchestration and experiment tracking (e.g., Kubeflow Pipelines, MLflow). - Experience with monitoring and observability tooling (e.g., Prometheus/Grafana, OpenTelemetry) and CI/CD for infra and ML workflows (e.g., GitHub Actions). At SpAItial, we are committed to creating a diverse and inclusive workplace. We welcome applications from people of all backgrounds, experiences, and perspectives. We are an equal opportunity employer and ensure all candidates are treated fairly throughout the recruitment process.

- Data systems and pipelines: Build end-to-end Python pipelines that turn third-party capture sources into clean, versioned training datasets — including scraping (e.g., Playwright) and preprocessing — and optimize the underlying storage at petabyte scale (object storage, fuse mounts, caching layers, shared filesystems, and relational / analytical / embedded metadata stores). - ML workflow orchestration and serving: Operate the systems researchers use to launch experiments, data jobs, and production endpoints — workflow engines (e.g., Kubeflow Pipelines, Airflow), GPU schedulers (e.g., Volcano, Slurm), experiment trackers (e.g., MLflow, Weights & Biases), and managed-inference platforms (e.g., Modal, Triton) — and maintain a launcher SDK for one-command runs. - Containerization and packaging: Ship workloads with Docker and Kubernetes; maintain IaC (Terraform) for the surfaces you own and CI/CD pipelines, including self-hosted GPU runners.

- Own and evolve the ML systems that enable training, evaluation, and serving of large foundation models — trainer, dataset loaders, checkpointing, and experiment orchestration code. - Distributed training enablement: Improve high-throughput training stacks (e.g., PyTorch DDP/FSDP, NCCL) for performance, stability, and reproducibility, including preemption-safe and sharded checkpointing. - Data systems and pipelines: Build end-to-end Python pipelines that turn third-party capture sources into clean, versioned training datasets — including scraping (e.g., Playwright) and preprocessing — and optimize the underlying storage at petabyte scale (object storage, fuse mounts, caching layers, shared filesystems, and relational / analytical / embedded metadata stores).

- 3+ years writing production-quality Python in a large, multi-author codebase, with strong SWE fundamentals (ML systems experience strongly preferred). - Hands-on with modern ML training stacks (PyTorch; DDP/FSDP or comparable); have personally debugged distributed jobs across many GPUs and nodes. - Have shipped non-trivial end-to-end data pipelines at scale — ingestion, transformation, validation, versioning, republish — ideally including real-world sources with rate limits, auth, or undocumented APIs.