Scaling AI training data collection to 250K+ contributors

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.

Twistag embedded three engineers directly into Defined.ai's product team for three months and built a React Native mobile application solving two problems: reaching contributors anywhere, and matching them to work they could actually do. The app lets contributors claim tasks offline, complete them with no network connection, and sync when they reconnect — the decision that unlocked markets where mobile data is expensive or unreliable.

On the backend Twistag built skill- and language-aware task routing, so a Portuguese-speaking transcription specialist sees Portuguese audio tasks first rather than Mandarin image labelling, and validation layers so quality held as volume grew. Twilio handles real-time notification when a high-value task drops into a contributor's region and language. The architecture is stateless, so scaling from 50,000 to 250,000 active contributors required no backend re-architecture. Twistag tested on low-bandwidth connections and older hardware, because the constraint was the infrastructure people connect through, not the phones they hold. Every constraint on this project was a product constraint wearing an infrastructure costume. The contributors the platform needed most were in markets where mobile data is expensive and connectivity unreliable, so offline claiming was not a refinement. A data-science bench asked to grow a contributor pool reaches for a matching algorithm; a product company asks what happens on a phone with two bars on a bus. Twistag ran the engagement as forward-deployed engineering — three engineers embedded inside Defined.ai's own product team for three months, rather than a delivery team behind a statement of work.

Stack: React Native, React, Twilio. Services: Product engineering, mobile engineering, AI-ready data platforms.

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 wasn't the phones people were using — it was the infrastructure they were connecting through.