Data Program Manager - World Models

This Is Growth; San Francisco (CA); On-site; USD 120,000 - 190,000; Full time

Job summary

Unknown AI lab in San Francisco is seeking a data collection operations lead to own end-to-end programs for multimodal data, including egocentric video, robotics, and sensors. You will translate researcher requests into scoped specs, design schemas, and set quality bars that teams can trust in deployment.

You will build QA infrastructure, run vendor management, capacity planning, and weekly reviews with researchers and engineers. This is a hands-on builder role in a fast-moving lab environment.

Qualifications

• Hands-on experience running physical-world data collection operations (egocentric video, robotics, AV data).
• Fluency in the model development lifecycle from data perspective (pre-training to eval).
• Ability to define taxonomies and quality criteria with researchers and engineers.
• Strong experience building QA systems, schemas, and data delivery SLAs.
• Willingness to work in the Bay Area or relocate.

Responsibilities

• Run field and lab data collection programs across multimodal data.
• Turn researcher requests into scoped specs with clear quality criteria.
• Maintain QA machinery with schema/timestamp/frame-drop checks on data.
• Manage collection vendors and capacity planning.

Skills

Data collection ops

Model lifecycle data

Cross-functional collaboration

Operational dashboards

Hands-on ownership

Relocation ready

XR/AR data

Schema ownership

Startup experience

Tools

QA dashboards

Job description

Our client is building general-purpose, causal, multimodal AI systems that simulate reality in real-time worlds you can step into and interact with. Their latest model holds the highest physics score among evaluated world models. Four model teams, six-plus data modalities, and one truth every frontier lab hits: the models are only as good as the data, and physical-world data doesn't come from scraping; it has to be collected, structured, and quality-gated by someone who's done it before.

The role:

You’ll own data collection and acquisition operations end to end the person who turns a researcher’s loose request into a scoped collection program with schemas, quality bars, and delivery dates, and who knows from experience where these programs break. You’ve run this loop before: validate on a small slice before committing volume, iterate the spec with researchers, scale the collection, hold back data for eval, and feed deployment gaps into the next request. You know why structured sequences beat continuous recordings, what an operational QA silo looks like, and what percentage capacity you hold in reserve before new requests get a red light.

What you’ll do

• Run field and lab data collection programs, capture hardware, trajectories, and scenario design across egocentric video, robotics, and multimodal data
• Turn ambiguous researcher requests into scoped specs: taxonomy agreed in advance, schema decisions made deliberately, quality criteria aligned before handoff
• Build the QA machinery: automated schema/timestamp/frame-drop checks on 100% of data, statistical human QA on top, pass-rate SLAs that mean something
• Manage collection vendors and operator teams’ capacity planning, throughput-per-operator economics, performance-based renewal decisions
• Run the forecasting and prioritization rhythm: live dashboards, capacity thresholds, weekly reviews with researchers and engineers
• Drive the retrospective loop what data moved which evals, at what cost, and what the next collection should do differently

What you bring

• Hands‑on experience running physical‑world data collection operations: egocentric video, robotics demonstration, sensor, motion, or AV data at meaningful scale
• Fluency in the model development lifecycle from the data side: you can explain what changes between pre‑training, fine‑tuning, and eval data without being prompted
• Direct working relationships with research scientists and ML engineers: you’ve negotiated taxonomies and quality criteria with the people consuming the data
• Operational spine: capacity models, QA systems and dashboards you built, with the numbers still in your head
• Comfort as a hands‑on owner at a 50‑person company; this is a builder seat, not a team‑management seat
• Bay Area‑based or relocating
• XR/AR data collection, human demonstration capture, or teleoperation data programs
• Experience where the collection hardware and the schema were yours to define
• Startup or 0→1 program build history

This is a chance to own the physical‑data engine at one of the most exciting AI labs in the world — where the collection programs you design directly move the frontier.