Unified training signal that shipped generative AI 5x faster
Problem
The team had spent many engineering cycles building its own training-data infrastructure for genAI products, which meant project delays and siloed AI development across disparate groups that couldn't find what data was available for ML use. Evaluating AI data quality was fragmented, lowering confidence in the models.
Solution
With Labelbox Catalog and Annotate and a variety of native editors, the company standardized how it produces and curates generative AI training signal — using metadata and embeddings to find unstructured data and built-in QA to keep signal quality high across teams.
Result
The company's genAI product teams saw significant gains in efficiency and speed using Labelbox's full platform — a 50% reduction in labeling operations time and a 5X increase in AI product deployment speed within just 8 months.
A Fortune 500 creative software company shipped generative AI across its products. Labelbox unified training-signal production and evaluation, cutting labeling time 50% and accelerating AI deployment 5x in 8 months.
The challenge
A Fortune 500 software company behind creative, marketing, and document-management products had been building AI into its cloud products for years. To ship generative AI across the portfolio, it needed one platform to produce and unify training signal. A major R&D division — working on video understanding via transcripts, joint vision and language, and document understanding — had been the first adopter. The team had spent heavy engineering cycles building its own training-data infrastructure, which meant delays and siloed generative AI work across groups that couldn't even find what data existed. Worse, evaluating AI data quality was fragmented, which lowered confidence in the models and stretched ROI timelines.
The approach
The company standardized on Labelbox across the division — one platform for ML and AI teams to collaborate with internal and external contributors, across text, PDF, image, and video. Labelbox Catalog used metadata and custom embeddings to filter unstructured data, saving hours a day on work that would have taken months of custom engineering. Labelbox Annotate gave the team a consistent process for producing high-quality signal, with built-in quality assurance that drove model-performance gains in NLP and computer vision — especially understanding the complex structure of PDF documents. Data science and product teams used the collaboration tools to define the data they wanted, translate requests into labeling instructions, and evolve ontologies as needs changed.
The outcome
The company cut labeling operations time by 50% and increased AI product deployment speed 5x within eight months. Its AI Assistant products — which understand PDF structure and content — shipped to production in 2023. With Labelbox's unified platform and labeling services, the company processes tens of thousands of PDFs through a dynamic queueing system that prioritizes the signal that improves generative AI outputs.
Where this goes
Shipping generative AI across a product line is a signal problem at scale. One platform to produce, curate, and evaluate training signal is what turns scattered AI projects into a portfolio.