Architect low-code AI solutions
What the exam expects you to know
Low-code on Google Cloud means assembling managed building blocks rather than writing training loops. The blocks you must recognise:
- Vertex AI Studio - prompt design, prompt management, model tuning (supervised fine-tuning and adapter tuning) and evaluation for Gemini models.
- Vertex AI Agent Builder (successor to Gen App Builder) - Vertex AI Search data stores for grounding (RAG) with citations, plus agents with tools/function calling. It is the answer whenever the requirement is "answer questions over our own documents with references".
- Dialogflow CX - deterministic, stateful conversational flows (IVR, slot filling, escalation). Pick it when the requirement mentions flows, intents and hand-off rather than free-form generation.
- Document AI - prebuilt processors, custom extractors, custom classifiers, Workbench, and human-in-the-loop review for forms and invoices.
- BigQuery ML and Looker Studio - SQL-first, low-code modelling and dashboards.
Traps
- Fine-tuning vs grounding. A question describing changing internal policy documents wants RAG over a Vertex AI Search data store, not a tuned model - tuning cannot keep facts fresh and gives no citations.
- Hidden non-functional requirement. A latency, data-residency, cost or "no data leaves the tenant" clause usually forces a different product than the obvious low-code one (e.g. VPC-SC + private connectivity, or BigQuery ML instead of exporting data).
- Low-code is not ungoverned. Every low-code tool still runs in a project, so IAM, CMEK, VPC Service Controls, quota and audit logging all apply.
Rule of thumb
Choose the least custom option that satisfies grounding and control: prebuilt API → Studio prompt + RAG → tuned model → fully custom training. Escalate only when a stated requirement blocks the previous rung.
Worked example: A bank wants an assistant over 200,000 internal policy PDFs that cites sources and is updated weekly. Answer: Vertex AI Search data store indexed from Cloud Storage, surfaced through an Agent Builder agent; re-index on change. Fine-tuning is wrong because citations and weekly freshness are required.