Datadog · Chan Zuckerberg Initiative · Salesforce (Internship) · Sprinklr · Microsoft (Internship)

The trust layer for Voxel51's consumption pricing. When customers pay by the token, the model call, and the gigabyte, opacity kills adoption – admins won't approve spend they can't see, explain, or bound. This page gives them all three: a full breakdown of what got consumed and by whom, plain-language accounting for every metered dimension so no line item is a mystery, and per-person limits with preset modes so a runaway agent stays a bounded problem instead of an end-of-month surprise.

Voxel51's first self-serve path into the product. You're in a real workspace with a curated dataset in under two minutes –  no SDK install, no Python, no bringing your own data first. Works across every data type the platform supports: images, video, point clouds, MCAP, LeRobot. The agent and embedding explorer are one click away the moment you land, so the first thing you do in the product is the thing you came to do

Voxel51's first native annotation system. Draw or fix labels by hand with SAM 3 in the loop, then train an agent on your own examples to label the rest of the dataset at scale. Project management, roles, and review pipelines live alongside it, so annotators, reviewers, and managers all work in one product where the data is also curated and the models are evaluated. Covers every label type visual AI needs including classifications, boxes, polygons, keypoints, semantic and instance segmentation, 3D cuboids, and video tracking across synchronized 2D and 3D.

shaped the 0→1 design of a new datadog product category, making enterprise-scale telemetry routing legible for mixed technical and non-technical teams.

brought pre-production observability into datadog: the mental model and core views for CI health across pipeline, test, and infrastructure signals.