Let’s Chat!

Scaling AI/ML Infrastructure for a Global Streaming Platform

Engagement: AI/ML Personalization & Data Solutions Program

Role: Senior Technical Product Manager, AI/ML Platforms & Data Engineering

Program Value: $650K in funded services across six mission-critical initiatives

The Challenge

Paramount+ needed to modernize the AI/ML backbone powering its content platform. The work spanned six distinct, technically demanding initiatives: automating artwork metadata tagging, building scalable model-serving infrastructure, standing up ephemeral training environments, detecting anomalies in real time, generating synthetic test data, and optimizing BigQuery costs and performance.

Each workstream touched a different part of the stack and a different set of stakeholders: engineering, data science, MLOps, DevOps, product, and curation/design. The teams were spread across the US, nearshore, and offshore teams. The risk was coordinating six interdependent efforts without losing velocity, alignment, or delivery discipline.

The Approach

I owned end-to-end product strategy and delivery across the program, defining requirements, roadmaps, and success metrics for each initiative while keeping every workstream aligned with a coherent whole. That meant operating as a technical thought partner to engineering on decisions such as model-serving architecture, pipeline design, and infrastructure trade-offs.

Automated artwork tagging platform. Directed the build of a metadata automation pipeline using Vertex AI/Gemini for extraction, paired with human-in-the-loop review workflows, a BigQuery metadata repository, observability tooling, and CI/CD automation, which reduced manual tagging effort while preserving quality control.

GKE-based inference gateway. Drove design and deployment of a containerized model-serving gateway supporting 3–10 large models concurrently, with autoscaling node pools, resource optimization policies, and secure CI/CD pipelines for safe, repeatable model rollouts.

Ephemeral training environments. Oversaw a GPU/TPU-based cluster orchestration system with incremental training-data detection, automated retraining workflows, versioned evaluation metrics, and a model promotion interface which made retraining a governed, repeatable process.

Real-time anomaly detection engine. Product-managed a Kafka/Datapflow ingestion pipeline feeding time-series ML models, with Vertex AI model registry governance, Superset dashboards, and alerting integrations to Slack, JIRA, and email gave teams early signals on issues before they became incidents.

Synthetic test data generation. Defined functional requirements and testing strategy for an AI-driven system that simulates test scenarios across 1–5 streaming and ETL schemas, integrated into CI/CD to catch data issues earlier in the pipeline.

Query optimization platform. Guided the development of a system that analyzes BigQuery audit logs using LLM-based evaluation to surface cost and latency optimization opportunities, delivered via dashboards and a developer-facing linting API.

Underpinning all six workstreams, I partnered with cloud architecture and DevOps to establish Terraform-driven infrastructure-as-code, GCP resource orchestration (GKE, Cloud Run, BigQuery, IAM, VPC), and enterprise-grade security standards including rigorous data governance, SSO/IAM compliance, and model versioning controls throughout.

Execution

Running six workstreams in parallel across distributed Agile teams required tight delivery discipline: milestone management, risk mitigation, alignment of dependencies across regions, and continuous stakeholder communication. I directed acceptance testing end-to-end to ensure each capability was production-ready before handoff.

The Impact

  • Six production AI/ML capabilities delivered across content operations, infrastructure, and data quality, from metadata automation to real-time anomaly detection.
  • Scalable model-serving infrastructure capable of running 3–10 large models concurrently with autoscaling and secure rollout pipelines.
  • Governed, repeatable ML operations retraining, promotion, and versioning processes that replaced ad hoc, manual workflows.
  • Cost and performance visibility into BigQuery usage, with developer-facing tooling to act on it.
  • Cross-regional program delivery at scale, with full compliance to enterprise data governance and security standards, across a $650K funded engagement.