Home / Customers

Delivery records from real projects

We prefer verifiable delivery numbers over logos. Customer names are partially anonymized unless authorization is in place; metrics come from platform delivery records.

Note: cases marked "illustrative" are structurally de-identified examples. Real customer identities and publishable figures are added once authorization is granted. We do not publish fabricated endorsements.
RoboticsIllustrative

Dual-arm grasping pipeline: 2,000 trajectory-labeled videos in 2 weeks

The customer needed ground truth with trajectories and keypoints for a kitchen-scene dual-arm grasping policy.
ChallengeScene setup and multi-camera synchronization
ApproachSite-based capture + teleoperation + per-batch QA
Result2,000 items / 14 days / 96.4% first-pass
DeliveryLeRobot + HDF5
Live CommerceIllustrative

Compliance evaluation set: better violation detection, lower review cost

The customer's moderation team applied inconsistent standards and had a high false-positive rate, with no reproducible benchmark.
ChallengeAmbiguous boundaries and drifting standards
ApproachStandard first + cross-labeling + arbitration
Result5,000 items / 8 days / Kappa 0.81
DeliveryJSON + CSV + evaluation report
Autonomous DrivingIllustrative

Long-tail roadside capture and 3D labeling

Model performance dropped on night and construction scenes, requiring targeted long-tail capture.
ChallengeScarce long-tail scenes and 3D GT precision
ApproachTargeted field crews + pre-labeling + multi-pass review
Result1,200 clips / 21 days / per-class recall passed
DeliveryKITTI + PCD
Content SafetyIllustrative

Consolidating scattered evaluation data into a standard dataset

The customer's evaluation data was scattered with no version control and could not be reused across downstream models.
ChallengeInconsistent metadata, no traceability
ApproachUnified metadata standard + version and lineage management
ResultReused by 3 downstream models, shorter delivery cycles
DeliveryMulti-format + version manifest

What customers say

Original feedback from acceptance and project reviews, published with permission.
"Long-tail capture was faster than we expected — and every batch came with a QA report, so issues were traceable to a specific batch."
— Data lead, autonomous driving lab
"The schema configurator saved a lot of back-and-forth. We imported our internal schema and the labeling standard finally became consistent."
— ML engineer, content safety team
"Delivering both LeRobot and HDF5 directly saved us a week of data wrangling."
— Co-founder, robotics team