Building an AI app: A low-code guide for small teams

Key Concepts

• Vibe Coding: A development methodology using AI-powered tools to generate code and implement features through natural language prompts rather than manual coding.
• Hybrid Search: A search technique combining traditional full-text search (keyword-based) with vector search (semantic/context-based) to improve accuracy.
• MCP (Model Context Protocol): An open standard that allows AI models to securely connect to data sources, tools, and APIs.
• Embeddings: Numerical representations of data (text, audio, etc.) that allow AI to understand semantic relationships (e.g., "big" and "large" are close in vector space).
• Agentic Workflows: Systems where AI agents don't just answer questions but perform actions (e.g., updating databases, confirming orders) via tools.
• Hill Climbing: An iterative process used by AI to refine database queries until they reach high accuracy.
• Parameterized Secure Views: A security mechanism that passes user identity to the database to ensure users only access data they are authorized to see.

---

1. Building AI Apps: The Low-Code Approach

The session emphasizes that small teams do not need massive infrastructure or specialized AI teams to modernize applications. The core strategy is to leverage existing databases (like Google Cloud’s AlloyDB) which now include built-in AI capabilities.

• The Four Pillars of Getting Started:
  1. Vibe Coding: Using AI assistants to write and refactor code.
  2. Hybrid Search: Implementing semantic search alongside traditional search.
  3. MCP Servers/Tools: Using standardized protocols to connect agents to data.
  4. Chatbot Integration: Adding conversational interfaces to existing business logic.

2. Step-by-Step: Implementing Semantic Search

The presenters demonstrated transforming a basic e-commerce "plant shop" into an AI-powered application:

1. Database Preparation: The agent inspects the schema and adds a vector column (e.g., description_vector).
2. Embedding Generation: Using models like text-embedding-005 to convert text into vectors.
3. Indexing: Creating a high-performance scan index on the vector column.
4. Frontend/Backend Update: The AI updates the API and UI to support a toggle between "Classic" (full-text) and "AI Mode" (semantic).
5. Result: The system can now understand that "pet friendly" is semantically equivalent to "harmless to dogs and cats," even if those specific words are absent from the product description.

3. Runtime Agents and MCP

The transition from "build-time" (developer assistance) to "runtime" (end-user experience) relies on MCP Servers.

• Workflow: The user asks a question $\rightarrow$ The Agent identifies the intent $\rightarrow$ The Agent selects a tool (e.g., get_applicable_promotions) $\rightarrow$ The tool queries the database $\rightarrow$ The Agent returns a natural language response.
• Managed vs. Open Source:
  • Managed Remote MCP Servers: Best for enterprise governance, IAM authentication, and auditing (e.g., BigQuery, AlloyDB).
  • MCP Toolbox for Databases: An open-source solution supporting 40+ data stores (MongoDB, Oracle, etc.) for local or multi-cloud deployments.

4. Real-World Application: CSA’s Modernization Journey

Diego Gamboa (CTO of CSA) shared a case study on modernizing a 45-year-old monolithic ERP system serving 11,000+ clients.

• The Challenge: A massive, shared codebase created an "innovation bottleneck" where simple features took months to deploy.
• The Solution:
  • Code Harvesting: An agent was built to document and store the logic of the legacy monolith.
  • Agent Framework: Developers shifted from writing code line-by-line to acting as "architects" who guide AI agents to build, test, and deploy microservices.
  • Architecture: The new system utilizes 11 Cloud Run microservices, 19 Pub/Sub topics, and 9 Postgres databases, all managed via Infrastructure as Code (IaC).

5. Notable Quotes

• "We have started moving from coding to genetic engineering." — Rahul Deshmukh, on the evolution of developer workflows.
• "Don't just take all of your corporate data, put it in one place and put AI on top of it... you need to separate this according to the people's roles." — Yoav, regarding security and data governance.
• "We are not doing agents that assist our teams; we are actually multiplying our teams." — Diego Gamboa, on the vision for AI-driven development.

6. Synthesis and Conclusion

The main takeaway is that modern databases are no longer just storage; they are the "heart" of AI applications. By utilizing MCP to bridge the gap between LLMs and structured data, and by adopting vibe coding to accelerate development, small teams can achieve high-level AI integration. The key to success is a phased approach: start with semantic search, move to agentic tools for specific tasks, and ensure strict security through parameterized views to maintain data integrity.