Data‑Centric AI and MLOps: A Case Study of Smart‑Cabin Applications in the Automotive Industry

Magic Data, founded in 2016, is a global AI data‑solution provider offering an intelligent annotation platform, AI datasets, and data‑collection services, and has served nearly 200 partners such as Microsoft, Qualcomm, NVIDIA, Alibaba, Baidu, and Tencent.

The concept of data‑centric AI contrasts with the traditional model‑centric approach by emphasizing systematic iteration of data inputs and labels rather than solely improving models. High‑quality data is essential and should be broad in coverage, multidimensional, timely, precise (balanced with cost), and compliant.

Data‑Centric MLOps is presented as the methodology to build efficient, data‑centric AI systems. It requires a multi‑party ecosystem and consists of scenario definition, data collection, data annotation, model training, and online deployment, forming a closed‑loop machine‑learning operation.

In the automotive smart‑cabin case, the workflow includes defining voice‑interaction scenarios, collecting audio data via Magic Data’s crowdsourcing platform, annotating data through three streams—ASR speech labeling, intent‑slot analysis for domain controllers, and image labeling for occupant monitoring—and training decentralized models that integrate with external or partner models.

Scaling the data from 3,000 to 30,000 hours yielded significant performance improvements: recognition rates increased by 5% for customer‑service dialogue, 9% for live‑social interaction, and 11% for in‑car navigation. Overall speech‑recognition error rates dropped 30%, with a 10% reduction in noisy environments.

The summary emphasizes that Data‑Centric AI focuses on data management rather than model tuning, continuous business‑loop updates are needed for sustained AI improvement, data challenges account for about 80% of AI work, and Data‑Centric MLOps lowers the technical barrier, enabling rapid intelligent transformation for businesses.