We build vetted contributor pools in India for AI labs — code & STEM annotation, evals, and RLHF preference data screened on your rubric, measured in every batch, and accountable to the founders. Built for AI teams in the US and Europe.
SCREENED ON YOUR RUBRIC → MEASURED IN EVERY BATCH → FOUNDER-ACCOUNTABLE
Models are only as good as the human judgment they learn from.
Most AI data comes from anonymous crowds graded on generic quizzes — hidden variance, unaccountable labels, acceptance tests that pass while behaviour degrades.
We run a different model.
Every Goldset contributor is screened on your rubric, scored against seeded gold in every batch, and traceable to a vetted expert under NDA.
We give you the scorecard — not just the labels. If a batch misses the agreed threshold, we redo it at our cost.
We're not a labelling vendor — we're the vetted human layer that turns expert judgment into trainable data, for language, code and — in pilot — physical AI.
Four workstreams, one standard of quality — scoped to your rubric, staffed by vetted specialists.
Code, maths and STEM judgment from screened engineers and PhDs — not a generic crowd.
We build rubrics, run them at scale, and surface the failure modes benchmarks miss.
Pairwise comparisons, rankings and rationales to train and stress-test reward models.
Native-tested data across India's major languages, for models going beyond English.
Vetted domain experts annotate professional-work video for embodied-AI and VLA training — causal event graphs, action-language captions, and 2D contact/keypoint labels, building on our own annotated multi-profession video work in India.
A task-specific certification, built from your spec, that contributors must pass (80%) before they touch production data.
Continuous QA with tracked inter-annotator agreement, so drift is caught in the batch — not by you, weeks later.
Every label traces to a vetted, NDA-bound contributor and a batch reviewer. No anonymous crowd.
You work directly with the people accountable for the outcome — and we redo off-threshold work at our cost.
We start behind an NDA, before any of your data changes hands.
Per-contributor NDA + IP assignment, a DPA, and non-circumvention on every engagement.
Access-controlled workspace, no subcontracting, activity logging, and a retention/deletion policy. GDPR & India DPDP-aware.
SOC 2 / ISO 27001. We share our current security posture on request.
Two founders on the ground, an expert pool sourced across India, and a focus on serving AI labs in the US and Europe. You work with the people who own the outcome — not an account queue.
Send us one hard task and a rubric. We'll vet a pool, label a sample, and show you measured quality — in 5–10 business days.