Who we are
AUGEN AI is headquartered in San Francisco and built by a Stanford team that has spent close to a decade building IoT, camera and video-intelligence products for physical industries. We've recruited 1,000+ customers from commercial businesses to large enterprises, and analyzed more than 10 billion minutes of video from real environments. We know what fails in the field: cameras that fog, networks that drop, operators who route around software.
What we’re doing
The AI wave that gave us ChatGPT and Claude is now expanding to the physical world. Robotics companies are teaching machines hands-on skills, and machines learn by watching skilled people. But understanding the world takes far more data than reading it, and almost none of what the world's cameras see is ever turned into anything a model can learn from.
We're closing that gap. We collect data from real US job sites and turn it into datasets that frontier labs train on. In the process, we help the businesses behind those job sites build a new revenue stream from their data.
Location: India, Chile or San Francisco Bay Area
Type: Contract to full time
Start: Immediately — first sample datasets ship within days of starting
The data you'll work with
- Real US job sites: manufacturing floors, warehouses and loading docks, active construction sites, and energy and utility plants. Not benchmarks, not staged demos.
- Every viewpoint of the same task: egocentric video from head-mounted stereo cameras, wrist cameras and IMUs, fused with fixed multi-camera exocentric views of the same scene.
- Video tied to how the work is supposed to be done: standard operating procedures mapped to worker video second by second, with task-completion scoring, plus machine telemetry from the equipment being operated.
- A ladder of products: raw time-synced footage for labs that label it themselves, and richly annotated datasets for labs that want them ready to train on. Your pipeline produces both.
What you'll own
- The end-to-end data pipeline, from raw upload to lab-ready delivery: ingest and cleaning, time sync across a head-mounted stereo camera, two wrist cameras and an IMU, face/PII anonymization, episode segmentation, QC, annotation, and export to LeRobot, RLDS and MCAP.
- You’ll build an automated, repeatable pipeline that turns each day's uploads into QC-scored, buyer-ready data.
- You’ll interface with data brokers and frontier labs to boost acceptance rates.
- You’ll make build vs. buy decisions by working with cutting edge vendors building solutions for the robotics stack.
- You'll manage the output of our annotation partners, judge outside processing vendors on their actual results, and feed hardware requirements back to the hardware owner.
- You'll build the high-value label layer on top: 3D hand pose, hand-object contact, and camera trajectory from stereo + IMU.
What excites you
- Data nobody else has. Internet video is running out of things to teach robots. You'll work on footage of skilled people doing real physical work in US industry, which no public dataset covers at scale.
- A genuinely hard fusion problem. Head camera, wrist cameras, IMU, fixed site cameras and machine sensors, all aligned into one record of a task. Getting that right is what makes the data trainable.
- Your pipeline is the product. What leaves your pipeline is what frontier labs buy. You'll see directly how your work shows up in the models they train.
- Ownership. You'll define the formats, the QC bar and the build-vs-buy calls, as one of the first engineers on a small team moving fast.
What you've done
- Recent PhD graduate, or 2–3 years building ML data infrastructure for multi-sensor systems (robotics, AR/VR, autonomous vehicles, or large-scale data curation).
- Built ML data infrastructure in production.
- Worked with multi-sensor systems: camera calibration and sync, visual-inertial SLAM, or 3D hand-pose estimation.
- Built a dataset or data pipeline that others actually trained on.
- Worked with production Python, comfortable moving terabytes and running GPU jobs, and willing to do hands-on QC.
- Bonus: egocentric video, UMI-style data, or robot-learning datasets.