What I build
The shape is practical: agents that call tools, touch business data, and can be deployed, tested, versioned, and reviewed.
Multi-agent architecture
Agents built on Google ADK and deployed through Agent Engine, with tools and MCP servers catalogued in Agent Registry. Behavior lives in versioned prompts, business rules as code, and tool contracts — not in a prompt someone edits in a console.
Google ADKAgent EngineAgent RegistryMCPtool calling
Vertex AI and Gemini in production
Model configuration, grounding, and native online evaluators on Vertex AI — and the boundary that matters in production: the model proposes, the system commits. Nothing irreversible happens on model output alone.
Vertex AIGeminionline evaluatorsgrounding
Evaluation engineering
Golden sets as code, an LLM judge calibrated case by case against hand-labeled turns, majority voting so a verdict is stable enough to block a release, hermetic per-case state, and parallel execution once cases are genuinely independent.
golden setsLLM judgecalibrationregression gateshermetic cases
Cloud delivery
Cloud Run services, Docker, a Terraform-managed estate, and CI/CD pipelines whose evaluation gates differ by environment — smoke on integration, full regression on QA — under trunk-based promotion.
Cloud RunDockerTerraformCI/CDAzure DevOps
Identity and enterprise security
IAM as the access model, Secret Manager for credentials, and PII boundaries enforced in telemetry — what the agent may do, what operators may see, and what never leaves the environment.
IAMSecret ManagerPII boundariesaccess control
Observability and cost
Traces and evaluator output exported to Cloud Monitoring, evaluation results as queryable data on BigQuery rather than logs read once, and cost per conversation tracked as a first-class figure.
Cloud MonitoringBigQuerytracingcost per conversation