How to build a financial analyst assistant with Vertex AI Studio & Gemini in under 10 minutes

Key Concepts

• Production-grade Financial Analyst Agent: An AI system designed to perform complex financial analysis tasks, capable of being deployed in a real-world, operational environment.
• Gemini: A powerful AI model developed by Google, used as the underlying engine for the agent.
• Vertex AI Studio: A platform that allows developers to build, test, and deploy AI models and applications without extensive setup.
• Express Mode: A feature within Vertex AI Studio that provides immediate access to AI tools without requiring a Google Cloud account, project selection, or billing information.
• Prompt Engineering: The process of designing and refining input text (prompts) to guide AI models to produce desired outputs.
• AI Agent for Prompt Optimization: An integrated AI assistant within Vertex AI Studio that helps developers improve their prompts by suggesting refinements and providing examples.
• JSON Output: A structured data format (JavaScript Object Notation) used for exchanging data, crucial for programmatic use by applications.
• Agent Builder: A tool within Vertex AI Studio for creating more complex AI agents that can chain multiple steps and utilize various tools.
• Tools (in Agent Builder): Reusable components that an agent can use, such as API calls (e.g., stock ticker API), database lookups, or pre-defined prompts.
• GCP Project Integration: The process of moving an AI project from Express Mode to a full Google Cloud Platform project for enterprise-grade features.
• Enterprise-grade Features: Capabilities like version control, collaboration, security, and governance essential for team-based development in a production environment.
• Read-only Sharing: A feature allowing users without direct access to the GCP project to view prompts, model settings, and sample outputs.

Building a Production-Grade Financial Analyst Agent in Under 10 Minutes

This video demonstrates how to rapidly build a production-grade financial analyst agent using Gemini powered by Vertex AI Studio, aiming to go from zero to a working prototype in under 10 minutes. The core challenge addressed is the typical friction developers face when starting AI projects, including lengthy setup, dependency management, and budget approvals.

Getting Started: Instant Access with Vertex AI Studio

The process begins by navigating to the Vertex AI Studio page. A key highlight is that no Google Cloud account login, project selection, or billing information is required to start prototyping. This "Express Mode" allows any developer to begin immediately and for free.

• Initial Exploration: Developers can explore the capabilities of different models using sample prompts, such as image generation with the nano banana model.

Developing the Financial Analyst Agent: From PDF to JSON

The primary goal is to build an agent that can process a long, complex quarterly earnings report PDF, extract key financial data, and structure it as clean JSON.

1. Initial Prompting and Trial-and-Error:

   • A developer starts by uploading a sample earnings report PDF.
   • A first-pass prompt is created to "extract the revenue, cost of goods sold, and the operating expenses from the uploaded PDF and list them."
   • Result: The output contains the information but is not in a parsable format, illustrating the common trial-and-error loop in prompt engineering. Developers often spend significant time rephrasing prompts to achieve the desired output structure.

2. AI-Assisted Prompt Optimization:

   • Vertex AI Studio integrates an AI agent to help refine prompts.
   • The developer uses a slash command (/prompt) and asks the agent to "help me refine this prompt to get a clean JSON output and with keys for revenue, cogs or cost of goods sold, and opex."
   • Key Improvement: The AI agent generates a more detailed prompt, including a persona, explicit instructions, and crucially, an example of the exact JSON structure required. This collaborative approach eliminates the guessing game.
   • Result: Running the optimized prompt against the same PDF yields perfect JSON output with correct keys and integer values, directly usable by visualization libraries like D3.

3. Automated Evaluation:

   • The "evaluate panel" allows developers to set up an "autorator" with a rubric.
   • Rules can be defined, such as "the output must be valid JSON" and "it must contain the revenue key."
   • When evaluations are run, the system automatically rates the model's responses against these criteria.

Transitioning to Production: From Prototype to Application

Once a satisfactory prompt and output are achieved, the focus shifts to integrating this into an application.

1. One-Click API Key Generation:

   • Developers can obtain an API key directly from the studio with a single click ("get API key").
   • This bypasses the need to navigate the GCP console, configure OAuth, or set up service accounts for immediate development.
   • For full production environments, this process ties into a service account.

2. Code Generation:

   • A "get code" button provides copy-pasteable code snippets, typically for VS Code.
   • Developers can integrate their API key and start coding their application.
   • A "build command" can reference the prompt within the context of an application, demonstrating how the work interacts with other application elements.

Scaling to Complex Agents and Team Collaboration

The platform supports building more sophisticated applications beyond simple prompt-based tasks.

1. Agent Builder for Multi-Step Processes:

   • For scenarios requiring fetching data from external APIs (e.g., stock data) and combining it with PDF analysis, a simple prompt is insufficient. This requires an agent.
   • The Agent Builder allows chaining multiple steps and integrating tools.
   • Tools can be:
     • Google Search
     • Database lookups
     • Custom APIs (e.g., stock ticker API)
   • The previously built prompt can be reused as a tool within a more complex agent. This demonstrates a direct scaling path from Express Mode prototyping to sophisticated multi-tool agents.

2. Enterprise-Grade Features for Team Projects:

   • Moving from Express Mode to a full GCP project unlocks enterprise-grade features for team collaboration.
   • Version History: Allows tracking changes to prompts, testing new ideas, and reverting to previous versions if errors occur. This is highly beneficial for tech leads.
   • Collaboration: Enables teams to work together on AI projects.
   • Governance and Security: Standard GCP security and governance practices are applied.

3. Read-Only Sharing for Stakeholders:

   • Users without GCP access (e.g., Product Managers, marketing teams) can be provided with a shareable link.
   • This link offers a read-only view of the prompt, model settings, and sample outputs, eliminating the need for excessive screenshots.

Conclusion and Key Takeaways

The demonstration successfully illustrates a streamlined workflow for building production-grade AI agents:

• Instant Access: Developers can start immediately in Express Mode without prior setup.
• AI-Assisted Prompting: The integrated AI agent significantly reduces the effort and guesswork in prompt engineering, leading to precise outputs like valid JSON.
• Rapid Integration: One-click API key generation and code snippets enable quick integration into applications.
• Scalability: Prompts can evolve into tools within sophisticated, multi-step agents.
• Team Collaboration and Governance: Seamless transition to full GCP projects provides essential enterprise features for secure, collaborative development.

The process moves from an idea to a production-grade, secure, and collaborative AI agent efficiently. The platform is designed to remove friction and accelerate AI development for a wide range of use cases, including financial analysis.