Session

How Databricks Apps Facilitated Model Retraining

Overview

ExperienceIn Person
TrackApplication Development
IndustryManufacturing
TechnologiesUnity Catalog, Databricks Apps, Lakebase
Skill LevelIntermediate
DOWNLOAD SESSION SLIDES

At Bridgestone, our mission is “serving society with superior quality.” As the world’s largest manufacturer of tire products, Bridgestone has a unique opportunity to leverage data and insights to improve manufacturing efficiency. Bridgestone’s Bandag Retread Solutions has over 65 years leading retread innovations.

 

In this talk, we will focus on how our data scientists used Databricks Apps to improve a computer vision model that facilitates decision-making during the tire retread process. We will show how in 3 weeks, a fully-fledged computer vision labeling application could be built using apps, saving our retread SMEs months of manual effort and thousands of dollars in a 3p software.

 

Our talk will also highlight the benefits of choosing Databricks apps to host our solution:

  • The ease of integration with existing Databricks resources
  • Out of box compliance with internal security standards
  • The speed in developing a front-end application, without the need of managing cloud infrastructure

Session Speakers

Gregory Islas

/Lead Data Scientist
Bridgestone Americas

Tim Hanchin

/Director
Bridgestone

Full Summary

How Bridgestone Turned SME Feedback Into a Model Retraining Engine With Databricks Apps

Bridgestone's mobility team needed a faster way to capture subject matter expert feedback on manufacturing images and turn that feedback into better vision models. Rather than maintain yet another external tool, the team built a fully integrated labeling app inside the Databricks Platform, shortening the loop from images to labels to retraining. For teams evaluating where to consolidate their data and AI workloads, the Databricks blog on building AI applications explains why a unified platform approach reduces operational overhead for lean engineering teams.

FAQ


For internal deployments, yes. Access flows through the existing identity and permissions model, either via the app's service principal or by passing the logged in user's token so actions run with that user's permissions.

Unity Catalog is optimized for analytics, not low latency transactional writes typical of web apps. Lakebase, built on Postgres, handled interactive saves smoothly, supported scale to zero for cost control, and could sync data back to Unity Catalog so monitoring dashboards continued to work.

The pattern is strongest for internal apps where users already have Databricks access. The speakers noted current limits around vertical scaling and external user access, and indicated that Databricks is working to expand these capabilities.

They used Databricks hosted LLMs, which kept prompts and context inside the governed Databricks environment. That allowed AI assisted development for the JavaScript frontend and other tasks while meeting enterprise security expectations.

Not yet. The app produces high quality labeled data used for retraining, and the team plans to automate the full loop so new labels trigger retraining and evaluation.