Case Study: FireworkTV boosts video recommendations and cuts ML costs with Provectus

FireworkTV reduces infrastructure costs and improves performance of its video recommendation system

FireworkTV, the decentralized short video network, needed to overhaul its machine learning infrastructure because its existing Lambda- and PyTorch-based setup was slowing productivity, increasing overhead, and limiting the quality and reliability of its real-time video recommendations. The company wanted a more automated, scalable AWS-based system to improve training and inference performance.

Provectus redesigned the ML stack using Amazon SageMaker, including migrated training, inference, and hosting pipelines, along with Amazon Elastic Inference and distributed training. With this new infrastructure, FireworkTV cut ML infrastructure costs by 2x, sped up inferences by 10x, and delivered the initial release in just four weeks, improving team productivity and real-time recommendation performance.

FireworkTV