What's happening and how this all works

Every AI model is trained on examples that people produced, checked or corrected. That work splits into collection (going out and capturing photos, audio or documents that don't exist online yet), annotation (labelling what's in that data so a model can learn from it) and evaluation (reading model output and judging it against a written standard). Jwuma pays for all three. None of it requires a technical background. It requires care and usually a language or a location that's hard to source.

Code-switching is when a speaker moves between languages inside a single sentence. Extremely common in Ghana, Kenya, the Philippines, India and much of the world and something most speech models handle badly. Tagging it means marking exactly where each language starts and stops. Only someone genuinely fluent in both languages can do it correctly, which is precisely why it's valuable and why it's priced above plain transcription.

A rubric is the written standard your work is judged against. The specific criteria, defined in advance. On Jwuma the rubric is taught to you in onboarding, frozen when the batch begins and used as the reference for any rejection. That freezing matters: it means the standard can't drift mid-project and leave you rejected for something that was acceptable last week.

Acceptance rate (accepted ÷ accepted + rejected), reliability (completed ÷ reserved) and average handling time. Corpshore runs a three-tier QA cascade on top of that. Reviewers themselves are measured on overrule rate, so a reviewer who consistently gets it wrong is visible to us and can be corrected. The measurement runs in both directions.

EXIF is the metadata your camera writes into a photo. When it was taken, on what device and often where. For training data that metadata is provenance: it's how a client proves a dataset is genuinely original field capture rather than scraped from the web. Jwuma preserves EXIF end to end, which is why capture tasks ask you to shoot in-app rather than upload from your gallery.

Two reasons. Clients sometimes require data to be handled only in specific jurisdictions for regulatory reasons and separately a small number of territories are blocked platform-wide for sanctions compliance. Either way you'll never see a listing you aren't eligible for, so restrictions cost you nothing but a listing you couldn't have taken anyway.