> Markdown version of [/jobs/ext/4628361-lead-engineer-ml-data-infrastructure-systems](https://www.wearedevelopers.com/jobs/ext/4628361-lead-engineer-ml-data-infrastructure-systems). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# Lead Engineer, ML Data Infrastructure & Systems

- **Company:** BUILD, Inc.
- **Location:** San Francisco, CA, United States
- **Experience:** Expert
- **Contract:** Permanent contract
- **Skills:** Airflow, Amazon Web Services, Cloud Computing, Encodings, Codecs, Data Files, Data Infrastructure, Extract Transform Load (ETL), Linux, Firmware, Python (Programming Language), Software Engineering, Management of Software Versions, Video Codec, Backend, Kubernetes, Deployment Automation
- **Published:** November 7, 2026
- **Apply:** https://www.juju.com/job/28_25197_37335304066c18222d0e86f7da487d24

## About the Role

* Pose, multi-camera, or other large media besides video
 * Cloud (AWS or GCP), orchestration (Kubernetes, Airflow, Temporal), or IaC
 * Dataset management or annotation tooling
 * You have shipped dataset delivery to external research or training customers

## Description

We're hiring a lead for the data platform: camera on a worker to training-ready datasets, and out to research customers. Collection is monocular 1920×1080p 30fps in the wild, targeting 100M hours. The roles we mean: Tesla Autopilot, Waymo, Cruise, Zoox, Nuro, Samsara, Verkada, Netflix encoding, YouTube ingest, Scale, Eventual/Daft. This is not a warehouse, analytics, or generic backend seat., * Own the data platform end-to-end: on-device capture, upload under flaky bandwidth, object storage, training-ready shards

 * Compression, codecs, and storage-tier trade-offs so 1080p30 hours stay cheap enough to keep collecting
 * Upload that survives bad networks: on-device buffering, batching, retries, a drop rate you can actually see
 * Object storage and training-shard formats. The hard problem is petabyte-scale media, not a warehouse
 * Own dataset packaging, versioning, and delivery to external research customers
 * Work with Shenzhen firmware so new devices speak one ingest contract, not a custom path per SKU
 * Make health, cost, and drop rate obvious as we add sites and countries, * You have owned a production media or sensor data path at real scale: object storage at petabyte scale, video codecs and compression, upload under flaky bandwidth, or training-shard / dataset formats
 * That kind of data path: Tesla Autopilot, Waymo, Cruise, Zoox, Nuro, Samsara, Verkada, Netflix encoding, YouTube ingest, Scale, or Eventual/Daft. Demo-scale ETL is not this job
 * Strong software engineering. Python and at least one systems language. Linux
 * You measure cost and throughput, not whether the demo uploaded
 * You want to scale in-the-wild physical-labor video, not run a generic data org