80%+ accuracy sustained a year after go-live across 5M+ transactions a month

An MLOps platform on Google Cloud Vertex AI that automates training, deployment, and monitoring for the client's audio quality models, retraining them in response to new data.

Outcome

5M+ transactions a month in live production, 80%+ accuracy held a year after go-live, drift monitored continuously, and retraining and deployment running as automated pipelines.

See the results

Talk to an amp'd specialist

At a glance

Client & industry

Global technology leader, running deep learning audio quality models in production

Engagement

MLOps platform design and implementation, with ongoing monitoring and support

Scale

5M+ transactions a month

Maps to below services

MLOps & Scalable ML Platforms

Where it started

Audio quality models that needed continuous retraining, with every step handled by hand.

Business problem

The audio quality models needed to retrain continuously for audio quality issues. The ML pipeline required a scalable platform, one that could adapt to changes in ML code and handle model versioning.

Technical barriers

The framework had to deploy, monitor, and maintain the models efficiently, with a consistent process for development, testing, and deployment. Model training, deployment, and uploads to the repository needed automating. CI/CD had to cover new code, model re-creation, and continuous deployment of new models, and monitoring, logging, and alerting had to track model effectiveness over time.

Agivant's role

MLOps architecture, Vertex AI pipeline engineering, CI/CD for ML, and ongoing model monitoring and support

What Agivant built

Agivant implemented an MLOps process that runs the ML workflow from development to production, using Google Cloud Vertex AI as a unified platform with serverless infrastructure and low-code and no-code implementation. Every step in the ML system is automated: code pushes trigger tested builds, pipelines extract features and train and evaluate models, and new versions move through the registry to serving with monitoring attached. Value in days and weeks.

ML development workspace

Data scientists and ML engineers develop in Vertex AI Workbench, with training code, tests, Dockerfiles, scripts, and configurations in GitHub.

CI on every code push

Each push triggers Cloud Build to run unit and integration tests, build Docker images, and compile the Kubeflow pipeline.

Container and artifact management

Container images for pipeline components go to Artifact Registry, and pipeline artifacts to Cloud Storage.

Automated training pipeline

Vertex AI Pipelines pull training and test data, extract features, train the model, and evaluate it.

Feature store

Features flow from BigQuery through Vertex AI Feature Store into training.

Model registry

Each trained model is registered with version control, audit trails, and accountability.

Model serving

Registered models deploy to Vertex AI Model Serving, powering the prediction service.

Model monitoring

Vertex AI Model Monitoring tracks performance, logs, and alerts, with continuous drift monitoring feeding the next retraining.

How it is built

Source and CI	Development	Pipelines	Data and features	Serving	Monitoring
GitHub, Cloud Build, Artifact Registry	Vertex AI Workbench	Vertex AI Pipelines, Kubeflow	BigQuery, Cloud Storage, Vertex AI Feature Store	Vertex AI Model Registry, Model Serving	Vertex AI Model Monitoring

Before

Manual training, deployment, and uploads, no consistent development-to-deployment process, and no ongoing view of model effectiveness.

After

Tested builds on every code push, automated retraining pipelines, a versioned registry, continuous deployment, and drift monitoring.

Proof in production

A data scientist pushes new training code, and Cloud Build runs the tests, builds the containers, and compiles the Kubeflow pipeline. Vertex AI trains and evaluates the model, registers the new version, and deploys it to serving. Monitoring tracks its performance and raises alerts on drift, feeding the next retraining run, and data scientists, ML engineers, IT operations, and business stakeholders all work from the same process.

Measured results

Models still accurate a year after go-live, with faster time to market, fewer resources needed for development and deployment, and every version traceable.

5M+

Transactions a month in live production

80%+

Accuracy sustained a year after go-live

Automated

Retraining and deployment pipelines

Continuous

Drift monitoring

Full

Audit trail and version control

The repeatable pattern

Reusable asset

The MLOps framework is an approved Agivant solution on Google Cloud, built on Vertex AI and customizable for other platforms, and it keeps the audio quality engine accurate in production. Every build amplifies the next.

Challenges overcome

Adapting the pipeline to frequent changes in ML code while handling model versioning, and retraining, adjusting, and redeploying models quickly in response to new data.

Enterprise movement

Foundational

Operational

Path to autonomous