Scaling AI training data collection to 250K+ contributors
Defined.ai's training-data marketplace ran on desktop and web, which shut out contributors in markets where a phone is the only connection. Twistag embedded three engineers in its product team and shipped a React Native app with offline tasks and language-aware routing in three months. Active contributors grew from 50,000 to more than 250,000 across 70-plus languages, on the original architecture.
- 250K+active contributors, up from 50K
- 70+languages covered
- 3 monthsconcept to production
- 10–14 daysclient data-collection cycles, down from weeks
- Client
- Defined.ai
- Industry
- Software & platforms
- Practices
- Data & Cloud, Product Engineering
- Timeline
- 3 months
- Stack
- React Native, React, Twilio
The contributors were on their phones.
Twistag built the mobile platform that took Defined.ai's AI training-data marketplace from 50,000 to more than 250,000 active contributors across 70-plus languages, from concept to production in three months. Defined.ai supplies labelled training data — image labelling, audio transcription, AI output validation — to enterprise clients including BMW and Mastercard, but its platform was desktop and web only, which excluded contributors in markets where a phone is the only connection. Twistag embedded three engineers in Defined.ai's product team and shipped a React Native app with offline task claiming and language-aware routing. Client data-collection cycles fell from multiple weeks to 10–14 days.
Training modern AI models requires large volumes of high-quality labelled data, and sourcing it across languages and geographies is fragmented, slow and expensive. Defined.ai's contributors were spread across more than 70 language markets, but the platform lived on desktop and web — which excluded mobile-only workers in exactly the regions the language coverage depended on. Without a mobile app, Defined.ai could not reach contributors at the moment they were free to work, and enterprise clients were getting slower turnaround on data collection than competitors offered. The company needed a production mobile app inside a single quarter.
Built for weak connections.
Offline by design.
Contributors claim tasks, complete them without a connection and sync when back online, which opened markets where mobile data is expensive or unreliable.
The right task for each person.
Skill- and language-aware routing shows a Portuguese transcription specialist Portuguese audio first. Twilio alerts contributors when a high-value task arrives in their region and language, and validation keeps quality steady as volume grows.
Ready to scale.
A stateless backend, tested on low bandwidth and older phones, took the platform from 50,000 to 250,000 contributors without re-architecture.
Embedded with the team.
Three Twistag engineers worked inside Defined.ai's own product team for the full three months.
Five times the contributors, same architecture.
The app shipped on schedule. Within six months Defined.ai went from 50,000 contributors to over 250,000 active participants across 70-plus languages. Downstream, clients including BMW and Mastercard moved from multi-week data-collection cycles to 10–14 day turnarounds. Language-specific task matching reduced contributor abandonment, and the offline queue produced higher completion rates and more predictable daily volume. The platform scaled 5x on its original architecture with no re-architecture, because the backend was stateless from the start. Defined.ai's enterprise clients include BMW, Mastercard, Accenture, Nuance and Voicebox.
The real constraint was the infrastructure people were connecting through.
- React Native
- React
- Twilio
Twistag built the mobile platform that took Defined.ai from 50,000 to more than 250,000 active contributors across 70-plus languages.
