Scaling AI training data collection to 250K+ contributors
Building a mobile platform that puts AI training work in the hands of a global crowd.
The challenge
Defined.ai faced a hard problem: training modern AI models requires massive amounts of high-quality labeled data, but sourcing that data at scale across languages and geographies is fragmented, slow, and expensive. The company had a pool of 250K+ potential contributors worldwide — people willing to label images, transcribe audio, and validate AI outputs — but no efficient way to route tasks to them. Contributors were scattered across 70+ language markets. The existing platform lived on desktop and web, which excluded mobile workers in regions where that's the only connection. Without a mobile app, Defined.ai couldn't reach contributors in their moment of availability, and Fortune 500 clients relying on the platform (BMW, Mastercard, Accenture, Nuance, Voicebox) were getting slower turnaround on training data collection than competitors offered.
The clock was ticking. Defined.ai needed a production mobile app in a single quarter, built by people who understood both mobile engineering and data workflows.
What we learned
| Mobile-only markets are invisible | When workers rely on mobile as their only connectivity, desktop-first platforms can't reach the labour pool. |
| Always-on assumes a privilege | Platforms that require constant connectivity exclude regions where bandwidth is intermittent and expensive. |
| Mismatch kills completion | When workers see tasks they can't do, abandonment climbs — skill-based routing is the floor. |
The solution
Twistag embedded three engineers directly into Defined.ai's product team for three months. We built a React Native mobile application that solved two critical problems: reaching contributors anywhere, and matching them to work they could actually do.
The app let contributors claim tasks offline, complete them without network connectivity, and sync results when they reconnected — critical for markets where mobile data is expensive or unreliable. On the backend, we implemented intelligent task matching: a contributor working in Portuguese with expertise in transcription would see Portuguese audio tasks first, not image labeling work in Mandarin. We built validation layers into the workflow so quality stayed high even as volume scaled.
We integrated Twilio for real-time notifications — when a high-value task dropped into a contributor's region and language, they found out immediately. The architecture was stateless, which meant scaling from 50K to 250K active contributors didn't require re-architecting the backend. We tested heavily on low-bandwidth connections and older hardware, because the real constraint wasn't the phones people were using — it was the infrastructure they were connecting through.
What this shaped
| Offline-first expands the market | Contributors claim and complete work offline, sync when connected — one decision unlocks unreachable workers. |
| Routing serves completion, not volume | Match tasks to contributors by language and expertise first — completion rate matters more than throughput. |
| Notification timing is engagement | When high-value work hits a contributor's region, they need to know immediately — real-time signal drives response. |
The impact
The mobile app shipped on schedule. Within six months, Defined.ai went from 50K contributors to over 250K active participants across 70+ languages. That velocity change cascaded downstream: clients like BMW and Mastercard went from multi-week data collection cycles to 10-14 day turnarounds. For a company whose entire value prop is speed of data delivery, that's the difference between winning a contract and losing it.
The offline-capable queue system became a core strength. Contributors could batch-claim 50 tasks during their lunch break over a weak WiFi connection, complete them on the bus without data, and sync when they got home. That behavioral freedom meant higher completion rates and more predictable daily volume — exactly what the supply-side metrics needed. Language-specific task matching reduced contributor abandonment because people weren't spending time on work they couldn't do.
What this proved
| Reliability beats sporadic volume | When workers can treat the platform as steady income, quality and consistency replace unpredictable, opportunistic output. |
| Architecture is the moat | Offline support and language matching aren't nice-to-have features — they're what make global scale possible at all. |
| New markets become operational | Once the platform works in one region with spotty connectivity, replicating to another is operations, not engineering. |
Technologies used
- React Native
- React
- Twilio
