MLOps & Model Serving
Package, serve, and monitor ML models on Vertex AI, Cloud Run, or Modal.
Production ML pipelines: training on Vertex AI, low-latency inference on Cloud Run or Modal, CI/CD, feature logging, drift alerts.
Problems I solve
- Models trained in notebooks, never shipped
- No CI/CD for retraining
- Silent drift degrading predictions
- Serving costs unpredictable
What you get
- Vertex AI / Modal training pipelines
- Autoscaling model endpoints
- Feature + prediction logging
- Drift detection and alerting
Use cases
Batch scoring pipeline
Nightly scoring at scale with lineage and versioning.
Real-time inference API
Low-latency endpoint with autoscaling and caching.
Feature store lite
Feature tables in BigQuery + online serving via Cloud Run.
Drift monitoring
Weekly drift + performance reports with alerts to Slack.
Examples I've shipped
Churn model serving
Outcome — p95 latency 45ms, 99.98% uptime.
GPU inference on Modal
Outcome — Serverless GPU serving with per-second billing.
How I work
- 1
Assess
Inventory models, data flows, and SLAs.
- 2
Package
Containerize training + serving with reproducible builds.
- 3
Deploy
Autoscaling endpoints with observability.
- 4
Monitor
Drift, performance, and cost dashboards.
Deliverables
- Training + serving pipelines
- CI/CD workflows
- Monitoring dashboards
- Runbook
Benefits
- Models actually in production
- Predictable serving costs
- Drift caught before customers notice
Frequently asked questions
Vertex AI or Modal?+
Vertex AI when you're already on GCP with BigQuery. Modal for GPU-heavy or bursty workloads with per-second billing.