# Reducing Rollbacks for Factory Computer Vision

> TechSupport Team supported PyTorch and MLflow deployments for VisionGuard’s factory computer vision system. Over five months, the client’s MLflow records show the rollback rate falling from 24% to 3%. The time from annotation to deployment fell from five weeks to nine days.

- Publisher: TechSupport Team
- Canonical page: https://techsupport-team.com/case-studies/industrial-vision-model-rollbacks/
- Updated: 2026-11-16

Client: VisionGuard · AI/ML & Data · German industrial computer vision (55 staff)

## Client and engagement facts

- **Client:** VisionGuard
- **Industry:** AI/ML & Data
- **Business context:** German industrial computer vision (55 staff)
- **Delivery model:** MLOps support
- **Application:** Industrial computer vision for factory QA
- **Supplied geography:** German
- **Company size in source:** 55 staff

## Application stack and tools

### Application stack

- Python
- PyTorch
- Kubernetes
- AWS

### Tools used

- MLflow
- Docker
- AWS billing
- Jira

Related services: [DevOps support](https://techsupport-team.com/services/devops-support/)

## Why Support Was Needed

Factory images differed from the staging data. The engineers assembled a reference test set from three months of production images and added a canary against factory conditions before model promotion.

## Technical Work

- Built a reference set from three months of production images.
- Added a canary check against factory images before promotion.
- Pinned model artifacts and GPU driver versions for each deployment.

## Representative Incident

A model trained on a bad annotation batch reached production in month three. The team rolled it back in 20 minutes and added a pre-deployment label check.

## Results and how they were measured

Table: Reported results and their measurement scope

| Metric | Baseline | Result | Units | Measurement window | Source and scope |
| --- | --- | --- | --- | --- | --- |
| Rollbacks | 24% | 3% | % | 5 months | Client MLflow |
| Inference cost | — | -37% | Result: % | 5 months | Client AWS |
| Annotation-to-deploy | 5 wk | 9 days | Baseline: wk; Result: days | 5 months | Client Jira |

## Evidence and Limitations

The comparison covers five months. Inference cost is reported as a relative change. The rollout incident remained part of the case’s account of the period.

## GPU Driver Version Control

The engineers pinned GPU driver versions per model in the Docker image and checked them at deployment time.

## Scope a Similar Engagement

Share how production images differ from staging and how your team tests and rolls back a model. [Request a support plan](https://techsupport-team.com/contact/) or explore [DevOps support](https://techsupport-team.com/services/devops-support/).

## Related pages

- [DevOps support](https://techsupport-team.com/services/devops-support/): The related service for this engagement.

## Contact

Request engineers from TechSupport Team or ask a question: https://techsupport-team.com/contact/
