AI Receptionist: My COMPLETE Workflow for Building Production AI Agents

Watch on YouTube: https://www.youtube.com/watch?v=9tinC8jG7FM

Channel: Owain Lewis

Introduction

The video presents a complete, repeatable workflow for building a production‑ready AI agent that could serve customers on a hotel website. It emphasizes the importance of upfront planning, using managed Google Cloud services, and leveraging AI‑assisted development tools to automate repetitive tasks. The goal is to give viewers a practical blueprint they can adapt to their own AI‑driven projects.

What the Video Covers

The walkthrough begins with gathering clear requirements and drafting a technical architecture document before any code is written. It shows how to bundle a static frontend and a backend API into a single Google Cloud Run service, select Gemini Flash as the inference model for its large context window, low latency, and low cost, and use AI agents such as Claude Code and Codex to generate tickets, write code, and run automated browser tests. The video also details setting up database backups, enabling observability, validating safety edge cases, and iterating on the agent using Google’s eval feature and LLM‑as‑a‑judge.

Step‑by‑Step Workflow

• Define requirements and create a requirements.md file that outlines functional and non‑functional goals for the AI agent.
• Draft an architecture.md document that describes how the frontend, API, model, and data stores will interact, noting that changes become costly later.
• Provision a Google Cloud Run service that bundles the static site and the API endpoint into a single managed deployment.
• Choose Gemini Flash as the model endpoint, taking advantage of its large context window, low latency, and inexpensive pricing.
• Use Claude Code or Codex to automatically generate Linear tickets (via MCP) and to scaffold code, reducing manual effort.
• Configure automated PostgrSQL backups with a retention policy, for example daily snapshots kept for seven days, to protect data and control storage growth.
• Activate Google Cloud Logging, Tracing, and Alerting to gain observability into request flows, errors, and performance metrics.
• Run the Google Eval suite, collect sample interactions, and employ an LLM‑as‑a‑judge to score safety and accuracy, then refine the system prompt stored in an external markdown file.
• Validate edge cases such as prompt injection and data leakage before promoting the agent to production.
• Break the effort into milestones — local development, testing, production deployment, and observability setup — to track progress and maintain focus.

Tools and Numbers Mentioned

• Google Cloud as the primary cloud provider.
• Cloud Run for hosting the combined frontend and API.
• Gemini Flash model selected for its large context window, low latency, and low cost.
• Claude Code and Codex used as AI agents to generate tickets, write code, and run browser‑based tests.
• Linear (via MCP) for ticket management and tracking.
• PostgrSQL as the relational database.
• Google Secret Manager for storing API keys and sensitive configuration.
• Google Logging & Tracing for observability.
• Agent CLI and Google Cloud CLI for infrastructure and agent management.
• Google Agent Development Kit (ADK) for building the agent.
• Google Eval feature for systematic evaluation and improvement.
• Example project value cited as $5,000 to illustrate scope.

Caveats and Hype Check

The video is produced in partnership with Google Cloud and uses a $5,000 hotel AI system as an illustrative example; this figure is not a guaranteed revenue outcome or result. The presenter stresses that spending significant time on evaluations and testing is critical for reliability, and that architectural changes become difficult once the system is deployed.

“If you’re building an AI system in production for real customers, you’re going to need all of these things.”

“Spending a lot of time on evals and testing is really really important.”

Concrete Next Actions

• Install the Google Cloud CLI and the Agent CLI to manage infrastructure and agent resources.
• Create a requirements.md file that captures all functional and non‑functional specifications.
• Draft an architecture.md document that outlines the Cloud Run service, model choice, data store, and security boundaries.
• Deploy a Cloud Run instance that bundles the static frontend and the backend API.
• Select Gemini Flash as the inference model for the agent, confirming its pricing and latency characteristics.
• Use Claude Code or Codex to generate Linear tickets and to scaffold initial code modules.
• Set up automated PostgrSQL backups with a defined retention policy, such as daily snapshots retained for seven days.
• Enable Google Cloud Logging, Tracing, and Alerting, and configure alerts for error rates and latency thresholds.
• Run the Google Eval suite, gather a representative set of user interactions, and use an LLM‑as‑a‑judge to evaluate safety and accuracy.
• Store the system prompt in an external markdown file to simplify versioning and updates.
• Perform targeted security testing for prompt injection and data leakage before moving to production.
• Organize work into milestones — local development, testing, production deployment, and observability — and track completion against each.