Monthly Archives: April 2025

Large language models (LLMs) have raised the bar for human-computer interaction where the expectation from users is that they can communicate with their applications through natural language. Beyond simple language understanding, real-world applications require managing complex workflows, connecting to external data, and coordinating multiple AI capabilities. Imagine scheduling a doctor’s appointment where an AI agent checks your calendar, accesses your provider’s system, verifies insurance, and confirms everything in one go—no more app-switching or hold times. In these real-world scenarios, agents can be a game changer, delivering more customized generative AI applications.

LLM agents serve as decision-making systems for application control flow. However, these systems face several operational challenges during scaling and development. The primary issues include tool selection inefficiency, where agents with access to numerous tools struggle with optimal tool selection and sequencing, context management limitations that prevent single agents from effectively managing increasingly complex contextual information, and specialization requirements as complex applications demand diverse expertise areas such as planning, research, and analysis. The solution lies in implementing a multi-agent architecture, which involves decomposing the main system into smaller, specialized agents that operate independently. Implementation options range from basic prompt-LLM combinations to sophisticated ReAct (Reasoning and Acting) agents, allowing for more efficient task distribution and specialized handling of different application components. This modular approach enhances system manageability and allows for better scaling of LLM-based applications while maintaining functional efficiency through specialized components.

This post demonstrates how to integrate open-source multi-agent framework, LangGraph, with Amazon Bedrock. It explains how to use LangGraph and Amazon Bedrock to build powerful, interactive multi-agent applications that use graph-based orchestration.

AWS has introduced a multi-agent collaboration capability for Amazon Bedrock Agents, enabling developers to build, deploy, and manage multiple AI agents working together on complex tasks. This feature allows for the creation of specialized agents that handle different aspects of a process, coordinated by a supervisor agent that breaks down requests, delegates tasks, and consolidates outputs. This approach improves task success rates, accuracy, and productivity, especially for complex, multi-step tasks.

Challenges with multi-agent systems

In a single-agent system, planning involves the LLM agent breaking down tasks into a sequence of small tasks, whereas a multi-agent system must have workflow management involving task distribution across multiple agents. Unlike single-agent environments, multi-agent systems require a coordination mechanism where each agent must maintain alignment with others while contributing to the overall objective. This introduces unique challenges in managing inter-agent dependencies, resource allocation, and synchronization, necessitating robust frameworks that maintain system-wide consistency while optimizing performance.

Memory management in AI systems differs between single-agent and multi-agent architectures. Single-agent systems use a three-tier structure: short-term conversational memory, long-term historical storage, and external data sources like Retrieval Augmented Generation (RAG). Multi-agent systems require more advanced frameworks to manage contextual data, track interactions, and synchronize historical records across agents. These systems must handle real-time interactions, context synchronization, and efficient data retrieval, necessitating careful design of memory hierarchies, access patterns, and inter-agent sharing.

Agent frameworks are essential for multi-agent systems because they provide the infrastructure for coordinating autonomous agents, managing communication and resources, and orchestrating workflows. Agent frameworks alleviate the need to build these complex components from scratch.

LangGraph, part of LangChain, orchestrates agentic workflows through a graph-based architecture that handles complex processes and maintains context across agent interactions. It uses supervisory control patterns and memory systems for coordination.

LangGraph Studio enhances development with graph visualization, execution monitoring, and runtime debugging capabilities. The integration of LangGraph with Amazon Bedrock empowers you to take advantage of the strengths of multiple agents seamlessly, fostering a collaborative environment that enhances the efficiency and effectiveness of LLM-based systems.

Understanding LangGraph and LangGraph Studio

LangGraph implements state machines and directed graphs for multi-agent orchestration. The framework provides fine-grained control over both the flow and state of your agent applications. LangGraph models agent workflows as graphs. You define the behavior of your agents using three key components:

• State – A shared data structure that represents the current snapshot of your application.
• Nodes – Python functions that encode the logic of your agents.
• Edges – Python functions that determine which Node to execute next based on the current state. They can be conditional branches or fixed transitions.

LangGraph implements a central persistence layer, enabling features that are common to most agent architectures, including:

• Memory – LangGraph persists arbitrary aspects of your application’s state, supporting memory of conversations and other updates within and across user interactions.
• Human-in-the-loop – Because state is checkpointed, execution can be interrupted and resumed, allowing for decisions, validation, and corrections at key stages through human input.

LangGraph Studio is an integrated development environment (IDE) specifically designed for AI agent development. It provides developers with powerful tools for visualization, real-time interaction, and debugging capabilities. The key features of LangGraph Studio are:

• Visual agent graphs – The IDE’s visualization tools allow developers to represent agent flows as intuitive graphic wheels, making it straightforward to understand and modify complex system architectures.
• Real-time debugging – The ability to interact with agents in real time and modify responses mid-execution creates a more dynamic development experience.
• Stateful architecture – Support for stateful and adaptive agents within a graph-based architecture enables more sophisticated behaviors and interactions.

The following screenshot shows the nodes, edges, and state of a typical LangGraph agent workflow as viewed in LangGraph Studio.

Figure 1: LangGraph Studio UI

In the preceding example, the state begins with __start__ and ends with __end__. The nodes for invoking the model and tools are defined by you and the edges tell you which paths can be followed by the workflow.

LangGraph Studio is available as a desktop application for MacOS users. Alternatively, you can run a local in-memory development server that can be used to connect a local LangGraph application with a web version of the studio.

Solution overview

This example demonstrates the supervisor agentic pattern, where a supervisor agent coordinates multiple specialized agents. Each agent maintains its own scratchpad while the supervisor orchestrates communication and delegates tasks based on agent capabilities. This distributed approach improves efficiency by allowing agents to focus on specific tasks while enabling parallel processing and system scalability.

Let’s walk through an example with the following user query: “Suggest a travel destination and search flight and hotel for me. I want to travel on 15-March-2025 for 5 days.” The workflow consists of the following steps:

1. The Supervisor Agent receives the initial query and breaks it down into sequential tasks:
   1. Destination recommendation required.
   2. Flight search needed for March 15, 2025.
   3. Hotel booking required for 5 days.
2. The Destination Agent begins its work by accessing the user’s stored profile. It searches its historical database, analyzing patterns from similar user profiles to recommend the destination. Then it passes the destination back to the Supervisor Agent.
3. The Supervisor Agent forwards the chosen destination to the Flight Agent, which searches available flights for the given date.
4. The Supervisor Agent activates the Hotel Agent, which searches for hotels in the destination city.
5. The Supervisor Agent compiles the recommendations into a comprehensive travel plan, presenting the user with a complete itinerary including destination rationale, flight options, and hotel suggestions.

The following figure shows a multi-agent workflow of how these agents connect to each other and which tools are involved with each agent.

Figure 2: Multi-agent workflow

Prerequisites

You will need the following prerequisites before you can proceed with this solution. For this post, we use the us-west-2 AWS Region. For details on available Regions, see Amazon Bedrock endpoints and quotas.

• A valid AWS account.
• An AWS Identity and Access Management (IAM) role in the account that has sufficient permissions to create the necessary resources.
• Access to Anthropic’s Claude 3 Sonnet and Claude 3.5 Sonnet in Amazon Bedrock. For instructions, see Access Amazon Bedrock foundation models.
• A LangGraph application up and running locally. For instructions, see Quickstart: Launch Local LangGraph Server.

Core components

Each agent is structured with two primary components:

• graph.py – This script defines the agent’s workflow and decision-making logic. It implements the LangGraph state machine for managing agent behavior and configures the communication flow between different components. For example:
  • The Flight Agent’s graph manages the flow between chat and tool operations.
  • The Hotel Agent’s graph handles conditional routing between search, booking, and modification operations.
  • The Supervisor Agent’s graph orchestrates the overall multi-agent workflow.
• tools.py – This script contains the concrete implementations of agent capabilities. It implements the business logic for each operation and handles data access and manipulation. It provides specific functionalities like:
  • Flight tools: search_flights, book_flights, change_flight_booking, cancel_flight_booking.
  • Hotel tools: suggest_hotels, book_hotels, change_hotel_booking, cancel_hotel_booking.

This separation between graph (workflow) and tools (implementation) allows for a clean architecture where the decision-making process is separate from the actual execution of tasks. The agents communicate through a state-based graph system implemented using LangGraph, where the Supervisor Agent directs the flow of information and tasks between the specialized agents.

To set up Amazon Bedrock with LangGraph, refer to the following GitHub repo. The high-level steps are as follows:

1. Install the required packages:

pip install boto3 langchain-aws

These packages are essential for AWS Bedrock integration:

• boto: AWS SDK for Python, handles AWS service communication
• langchain-aws: Provides LangChain integrations for AWS services

2. Import the modules:

from langchain_aws import ChatBedrockConverse 
from langchain_aws import ChatBedrock

3. Create an LLM object:

bedrock_client = boto3.client("bedrock-runtime", region_name="<region_name>")
llm = ChatBedrockConverse(
        model="anthropic.claude-3-haiku-20240307-v1:0",
        temperature=0,
        max_tokens=None,
        client=bedrock_client,
        # other params...
    )

LangGraph Studio configuration

This project uses a langgraph.json configuration file to define the application structure and dependencies. This file is essential for LangGraph Studio to understand how to run and visualize your agent graphs.

{
"dependencies": [
"boto3>=1.35.87",
"langchain-aws>=0.2.10",
"."
],
"graphs": {
"supervisor": "./src/supervisor_agent/graph.py:graph",
"flight": "./src/flight_agent/graph.py:graph",
"hotel": "./src/hotel_agent/graph.py:graph"
},
"env": "./.env"
}

LangGraph Studio uses this file to build and visualize the agent workflows, allowing you to monitor and debug the multi-agent interactions in real time.

Testing and debugging

You’re now ready to test the multi-agent travel assistant. You can start the graph using the langgraph dev command. It will start the LangGraph API server in development mode with hot reloading and debugging capabilities. As shown in the following screenshot, the interface provides a straightforward way to select which graph you want to test through the dropdown menu at the top left. The Manage Configuration button at the bottom lets you set up specific testing parameters before you begin. This development environment provides everything you need to thoroughly test and debug your multi-agent system with real-time feedback and monitoring capabilities.

Figure 3: LangGraph studio with Destination Agent recommendation

LangGraph Studio offers flexible configuration management through its intuitive interface. As shown in the following screenshot, you can create and manage multiple configuration versions (v1, v2, v3) for your graph execution. For example, in this scenario, we want to use user_id to fetch historic use information. This versioning system makes it simple to track and switch between different test configurations while debugging your multi-agent system.

Figure 4: Runnable configuration details

In the preceding example, we set up the user_id that tools can use to retrieve history or other details.

Let’s test the Planner Agent. This agent has the compare_and_recommend_destination tool, which can check past travel data and recommend travel destinations based on the user profile. We use user_id in the configuration so that can it be used by the tool.

LangGraph has concept of checkpoint memory that is managed using a thread. The following screenshot shows that you can quickly manage threads in LangGraph Studio.

Figure 5: View graph state in the thread

In this example, destination_agent is using a tool; you can also check the tool’s output. Similarly, you can test flight_agent and hotel_agent to verify each agent.

When all the agents are working well, you’re ready to test the full workflow. You can evaluate the state a verify input and output of each agent.

The following screenshot shows the full view of the Supervisor Agent with its sub-agents.

Figure 6: Supervisor Agent with complete workflow

Considerations

Multi-agent architectures must consider agent coordination, state management, communication, output consolidation, and guardrails, maintaining processing context, error handling, and orchestration. Graph-based architectures offer significant advantages over linear pipelines, enabling complex workflows with nonlinear communication patterns and clearer system visualization. These structures allow for dynamic pathways and adaptive communication, ideal for large-scale deployments with simultaneous agent interactions. They excel in parallel processing and resource allocation but require sophisticated setup and might demand higher computational resources. Implementing these systems necessitates careful planning of system topology, robust monitoring, and well-designed fallback mechanisms for failed interactions.

When implementing multi-agent architectures in your organization, it’s crucial to align with your company’s established generative AI operations and governance frameworks. Prior to deployment, verify alignment with your organization’s AI safety protocols, data handling policies, and model deployment guidelines. Although this architectural pattern offers significant benefits, its implementation should be tailored to fit within your organization’s specific AI governance structure and risk management frameworks.

Clean up

Delete any IAM roles and policies created specifically for this post. Delete the local copy of this post’s code. If you no longer need access to an Amazon Bedrock FM, you can remove access from it. For instructions, see Add or remove access to Amazon Bedrock foundation models

Conclusion

The integration of LangGraph with Amazon Bedrock significantly advances multi-agent system development by providing a robust framework for sophisticated AI applications. This combination uses LangGraph’s orchestration capabilities and FMs in Amazon Bedrock to create scalable, efficient systems. It addresses challenges in multi-agent architectures through state management, agent coordination, and workflow orchestration, offering features like memory management, error handling, and human-in-the-loop capabilities. LangGraph Studio’s visualization and debugging tools enable efficient design and maintenance of complex agent interactions. This integration offers a powerful foundation for next-generation multi-agent systems, providing effective workflow handling, context maintenance, reliable results, and optimal resource utilization.

For the example code and demonstration discussed in this post, refer to the accompanying GitHub repository. You can also refer to the following GitHub repo for Amazon Bedrock multi-agent collaboration code samples.

About the Authors

Jagdeep Singh Soni is a Senior Partner Solutions Architect at AWS based in the Netherlands. He uses his passion for generative AI to help customers and partners build generative AI applications using AWS services. Jagdeep has 15 years of experience in innovation, experience engineering, digital transformation, cloud architecture, and ML applications.

Ajeet Tewari is a Senior Solutions Architect for Amazon Web Services. He works with enterprise customers to help them navigate their journey to AWS. His specialties include architecting and implementing scalable OLTP systems and leading strategic AWS initiatives.

Rupinder Grewal is a Senior AI/ML Specialist Solutions Architect with AWS. He currently focuses on serving of models and MLOps on Amazon SageMaker. Prior to this role, he worked as a Machine Learning Engineer building and hosting models. Outside of work, he enjoys playing tennis and biking on mountain trails.

Read More

Generative AI enables us to accomplish more in less time. Text-to-SQL empowers people to explore data and draw insights using natural language, without requiring specialized database knowledge. Amazon Web Services (AWS) has helped many customers connect this text-to-SQL capability with their own data, which means more employees can generate insights. In this process, we discovered that a different approach is needed in enterprise environments where there are over 100 tables, each with dozens of columns. We also learned that robust error handling is critical when errors occur in the generated SQL query based on users’ questions.

This post demonstrates how enterprises can implement a scalable agentic text-to-SQL solution using Amazon Bedrock Agents, with advanced error-handling tools and automated schema discovery to enhance database query efficiency. Our agent-based solution offers two key strengths:

1. Automated scalable schema discovery – The schema and table metadata can be dynamically updated to generate SQL when the initial attempt to execute the query fails. This is important for enterprise customers who have a lot of tables and columns and many queries patterns.
2. Automated error handling – The error message is directly fed back to the agent to improve the success rate of running queries.

You’ll find that these features help you tackle enterprise-scale database challenges while making your text-to-SQL experience more robust and efficient.

Use case

An agentic text-to-SQL solution can benefit enterprises with complex data structures. In this post, to understand the mechanics and benefits of the agentic text-to-SQL solution in a complex enterprise environment, imagine you’re a business analyst on the risk management team in a bank. You need to answer questions such as “Find all transactions that occurred in the United States and were flagged as fraudulent, along with the device information used for those transactions,” or “Retrieve all transactions for John Doe that occurred between January 1, 2023, and December 31, 2023, including fraud flags and merchant details.” For this, there are dozens—or sometimes hundreds—of tables that you need to not only be aware but also craft complex JOIN queries. The following diagram illustrates a sample table schema that might be needed for fraud investigations.

The key pain points of implementing a text-to-SQL solution in this complex environment include the following, but aren’t limited to:

1. The amount of table information and schema will get excessive, which will entail manual updates on the prompts and limit its scale.
2. As a result, the solution might require additional validation, impacting the quality and performance of generating SQL.

Now, consider our solution and how it addresses these problems.

Solution overview

Amazon Bedrock Agents seamlessly manages the entire process from question interpretation to query execution and result interpretation, without manual intervention. It seamlessly incorporates multiple tools, and the agent analyzes and responds to unexpected results. When queries fail, the agent autonomously analyzes error messages, modifies queries, and retries—a key benefit over static systems.

As of December 2024, the Amazon Bedrock with structured data feature provides built-in support for Amazon Redshift, offering seamless text-to-SQL capabilities without custom implementation. This is recommended as the primary solution for Amazon Redshift users.

Here are the capabilities that this solution offers:

1. Executing text-to-SQL with autonomous troubleshooting:
   1. The agent can interpret natural language questions and convert them into SQL queries. It then executes these queries against an Amazon Athena database and returns the results.
   2. If a query execution fails, the agent can analyze the error messages returned by AWS Lambda and automatically retries the modified query when appropriate.
2. Dynamic schema discovery
   1. Listing tables – The agent can provide a comprehensive list of the tables in the fraud detection database. This helps users understand the available data structures.
   2. Describing table schemas – Users can request detailed information about the schema of specific tables. The agent will provide column names, data types, and associated comments, giving users a clear understanding of the data structure.

The solution uses direct database tools for schema discovery instead of vector store–based retrieval or static schema definitions. This approach provides complete accuracy with lower operational overhead because it doesn’t require a synchronization mechanism and continually reflects the current database structure. Direct schema access through tools is more maintainable than hardcoded approaches that require manual updates, and it provides better performance and cost-efficiency through real-time database interaction.

The workflow is as follows:

1. A user asks questions to Amazon Bedrock Agents.
2. To serve the user’s questions, the agent determines the appropriate action to invoke:
   1. To execute the generated query with confidence, the agent will invoke the athena-query
   2. To confirm the database schema first, the agent will invoke the athena-schema-reader tool:
      • Retrieve a list of available tables using its /list_tables endpoint.
      • Obtain the specific schema of a certain table using its /describe_table endpoint.
   3. The Lambda function sends the query to Athena to execute.
   4. Athena queries the data from the Amazon Simple Storage Service (Amazon S3) data bucket and stores the query results in the S3 output bucket.
   5. The Lambda function retrieves and processes the results. If an error occurs:
      • The Lambda function captures and formats the error message for the agent to understand.
      • The error message is returned to Amazon Bedrock Agents.
      • The agent analyzes the error message and tries to resolve it. To retry with the modified query, the agent may repeat steps 2–5.
   6. The agent formats and presents the final responses to the user.

The following architecture diagram shows this workflow.

Implementation walkthrough

To implement the solution, use the instructions in the following sections.

Intelligent error handling

Our agentic text-to-SQL solution implements practical error handling that helps agents understand and recover from issues. By structuring errors with consistent elements, returning nonbreaking errors where possible, and providing contextual hints, the system enables agents to self-correct and continue their reasoning process.

Agent instructions

Consider the key prompt components that make this solution unique. Intelligent error handling helps automate troubleshooting and refine the query by letting the agent understand the type of errors and what to do when error happens:

Execution and Error Handling:
   - Execute the query via the /athena_query endpoint
   - If the execution fails, carefully analyze the error message and hint provided by the Lambda function
   - Based on the error type received from the Lambda function, take appropriate action:
   - After identifying the issue based on the error message and hint:
     1. Modify your query or API request to address the specific problem
     2. If needed, use schema discovery tools (/list_tables, /describe_table) to gather updated information
     3. Reconstruct the query with the necessary corrections
     4. Retry the execution with the modified query or request

The prompt gives guidance on how to approach the errors. It also states that the error types and hints will be provided by Lambda. In the next section, we explain how Lambda processes the errors and passes them to the agent.

Implementation details

Here are some key examples from our error handling system:

ERROR_MESSAGES = {
    'QUERY_EXECUTION_FAILED': {
        'message': 'Failed to execute query',
        'hint': 'Please use fully qualified table names. Example: SELECT * FROM fraud_data.customers LIMIT 1'
    },
    'QUERY_RESULT_ERROR': {
        'message': 'Error occurred while getting query results',
        'hint': 'Check if the tables and columns in your query exist and you have proper permissions. Examples: "customers", "transactions", or "devices".'
    },
    'MISSING_QUERY': {
        'message': 'Query is required',
        'hint': 'No query was provided. Please provide a SQL query to execute'
    }
}

def create_query_response(query_result, status_code=200):
    if query_result.get('error'):
        error_info = ERROR_MESSAGES.get(query_result['error'])
        return {
            'error': query_result['error'],
            'message': error_info['message'],
            'hint': error_info['hint']
        }
    return query_result

These error types cover the main scenarios in text-to-SQL interactions:

1. Query execution failures – Handles syntax errors and table reference issues, guiding the agent to use the correct table names and SQL syntax
2. Result retrieval issues – Addresses permission problems and invalid column references, helping the agent verify the schema and access rights
3. API validation – Verifies that basic requirements are met before query execution, minimizing unnecessary API calls

Each error type includes both an explanatory message and an actionable hint, enabling the agent to take appropriate corrective steps. This implementation shows how straightforward it can be to enable intelligent error handling; instead of handling errors traditionally within Lambda, we return structured error messages that the agent can understand and act upon.

Dynamic schema discovery

The schema discovery is pivotal to keeping Amazon Bedrock Agents consuming the most recent and relevant schema information.

Agent instructions

Instead of hardcoded database schema information, we allow the agent to discover the database schema dynamically. We’ve created two API endpoints for this purpose:

Schema Discovery: 
    - Use /list_tables endpoint to identify available tables in the database 
    - Use /describe_table endpoint to get detailed schema information for specific tables 
    - Always use the most recent and relevant table schemas, as the database structure may change frequently 
    - Before constructing queries, ensure you have up-to-date schema information

Implementation details

Based on the agent instructions, the agent will invoke the appropriate API endpoint.

The /list_tables endpoint lists the tables in a specified database. This is particularly useful when you have multiple databases or frequently add new tables:

@app.post("/list_tables", description="Retrieve a list of all tables in the specified database")
def list_tables(event, database_name):
    query = f"SHOW TABLES IN {database_name}"
    result = execute_and_get_results(query, s3_output)
    if isinstance(result, dict) and 'error' in result:
        return create_api_response(event, 400, get_error_response('QUERY_RESULT_ERROR'))
    return create_api_response(event, 200, result)

The /describe_table endpoint reads a specific table’s schema with details. We use the “DESCRIBE” command, which includes column comments along with other schema details. These comments help the agent better understand the meaning of the individual columns:

@app.post("/describe_table", description="Retrieve the schema information of a specific table")
def describe_table(event, database_name, table_name):
    query = f"DESCRIBE {database_name}.{table_name}"
    result = execute_and_get_results(query, s3_output)
    
    if isinstance(result, dict) and 'error' in result:
        return create_api_response(event, 400, get_error_response('QUERY_RESULT_ERROR'))
    
    formatted_result = {
        "table_name": table_name,
        "database": database_name,
        "columns": result
    }
    return create_api_response(event, 200, formatted_result)

When implementing a dynamic schema reader, consider including comprehensive column descriptions to enhance the agent’s understanding of the data model.

These endpoints enable the agent to maintain an up-to-date understanding of the database structure, improving its ability to generate accurate queries and adapt to changes in the schema.

Demonstration

You might not experience the exact same response with the presented screenshot due to the indeterministic nature of large language models (LLMs).

The solution is available for you to deploy in your environment with sample data. Clone the repository from this GitHub link and follow the README guidance. After you deploy the two stacks—AwsText2Sql-DbStack and AwsText2Sql-AgentStack—follow these steps to put the solution in action:

1. Go to Amazon Bedrock and select Agents.
2. Select AwsText-to-SQL-AgentStack-DynamicAgent and test by asking questions in the Test window on the right.
3. Example interactions:
   • Which demographic groups or industries are most frequently targeted by fraudsters? Present aggregated data.
   • What specific methods or techniques are commonly used by perpetrators in the reported fraud cases?
   • What patterns or trends can we identify in the timing and location of fraud incidents?
   • Show the details of customers who have made transactions with merchants located in Denver.
   • Provide a list of all merchants along with the total number of transactions they’ve processed and the number of those transactions that were flagged as fraudulent.
   • List the top five customers based on the highest transaction amounts they’ve made.

4. Choose Show trace and examine each step to understand what tools are used and the agent’s rationale for approaching your question, as shown in the following screenshot.

5. (Optional) You can test the Amazon Bedrock Agents code interpreter by enabling it in Agent settings. Follow the instructions at Enable code interpretation in Amazon Bedrock and ask the agent “Create a bar chart showing the top three cities that have the most fraud cases.”

Best practices

Building on our discussion of dynamic schema discovery and intelligent error handling, here are key practices to optimize your agentic text-to-SQL solution:

1. Use dynamic schema discovery and error handling – Use endpoints such as /list_tables and /describe_table to allow the agent to dynamically adapt to your database structure. Implement comprehensive error handling as demonstrated earlier, enabling the agent to interpret and respond to various error types effectively.
2. Balance static and dynamic information – Although dynamic discovery is powerful, consider including crucial, stable information in the prompt. This might include database names, key table relationships, or frequently used tables that rarely change. Striking this balance can improve performance without sacrificing flexibility.
3. Tailoring to your environment – We designed the sample to always invoke /list_tables and /describe_table, and your implementation might need adjustments. Consider your specific database engine’s capabilities and limitations. You might need to provide additional context beyond only column comments. Think about including database descriptions, table relationships, or common query patterns. The key is to give your agent as much relevant information as possible about your data model and business context, whether through extended metadata, custom endpoints, or detailed instructions.
4. Implement robust data protection – Although our solution uses Athena, which inherently doesn’t support write operations, it’s crucial to consider data protection in your specific environment. Start with clear instructions in the prompt (for example, “read-only operations only”), and consider additional layers such as Amazon Bedrock Guardrails or an LLM-based review system to make sure that generated queries align with your security policies.
5. Implement layered authorization – To enhance data privacy when using Amazon Bedrock Agents, you can use services such as Amazon Verified Permissions to validate user access before the agent processes sensitive data. Pass user identity information, such as a JWT token, to the agent and its associated Lambda function, enabling fine-grained authorization checks against pre-built policies. By enforcing access control at the application level based on the Verified Permissions decision, you can mitigate unintended data disclosure and maintain strong data isolation. To learn more, refer to Enhancing data privacy with layered authorization for Amazon Bedrock Agents in the AWS Security Blog.
6. Identify the best orchestration strategy for your agent – Amazon Bedrock provides you with an option to customize your agent’s orchestration strategy. Custom orchestration gives you full control of how you want your agents to handle multistep tasks, make decisions, and execute workflows.

By implementing these practices, you can create a text-to-SQL solution that not only uses the full potential of AI agents, it also maintains the security and integrity of your data systems.

Conclusion

In conclusion, the implementation of a scalable agentic text-to-SQL solution using AWS services offers significant advantages for enterprise workloads. By using automated schema discovery and robust error handling, organizations can efficiently manage complex databases with numerous tables and columns. The agent-based approach promotes dynamic query generation and refinement, leading to higher success rates in data querying. We’d like to invite you to try this solution out today! Visit GitHub to dive deeper into the details of the solution, and follow the deployment guide to test in your AWS account.

About the Authors

Jimin Kim is a Prototyping Architect on the AWS Prototyping and Cloud Engineering (PACE) team, based in Los Angeles. With specialties in Generative AI and SaaS, she loves helping her customers succeed in their business. Outside of work, she cherishes moments with her wife and three adorable calico cats.

Jiwon Yeom is a Solutions Architect at AWS, based in New York City. She focuses on Generative AI in the financial services industry and is passionate about helping customers build scalable, secure, and human-centered AI solutions. Outside of work, she enjoys writing, and exploring hidden bookstores.

Read More

Many organizations rely on multiple third-party applications and services for different aspects of their operations, such as scheduling, HR management, financial data, customer relationship management (CRM) systems, and more. However, these systems often exist in silos, requiring users to manually navigate different interfaces, switch between environments, and perform repetitive tasks, which can be time-consuming and inefficient.

Moreover, while many enterprise systems are equipped with APIs for integration, users often lack the technical expertise to interact with these APIs directly. As a result, organizations need an intuitive and seamless way to query data and perform actions across these applications using natural language, without requiring specialized knowledge of each system or its APIs.

To address the challenge of integrating multiple third-party applications into a unified, natural language-driven interface, users can use plugins for Amazon Q Business. Plugins provide a way to bridge the gap between complex, siloed enterprise applications in a user-friendly interfacing empowering users to take action across systems with easy. Amazon Q Business supports multiple enterprise systems with pre-built plugins, as well as custom plugins, that users can use to integrate a variety of enterprise systems with Amazon Q Business applications.

Solution overview

In this post, we demonstrate how you can use custom plugins for Amazon Q Business to build a chatbot that can interact with multiple APIs using natural language prompts. We showcase how to build an AIOps chatbot that enables users to interact with their AWS infrastructure through natural language queries and commands. The chatbot is capable of handling tasks such as querying the data about Amazon Elastic Compute Cloud (Amazon EC2) ports and Amazon Simple Storage Service (Amazon S3) buckets access settings. For example, users can ask the chatbot questions like “Which EC2 instances have port 3389 open?” or request actions such as “Please close public access for S3 buckets.”

By integrating other AWS services with Amazon Q using OpenAPI schemas, the chatbot can not only retrieve real-time information (such as checking which S3 buckets have public access), but also take corrective actions (such as closing open ports or public access) in response to user commands. This solution reduces manual intervention and simplifies complex cloud operations by enabling IT teams to manage infrastructure through natural language interactions. The chatbot will streamline operational tasks, reduce the need for switching between different tools, and improve the efficiency of IT and operations teams by allowing them to interact with complex systems using simple, intuitive language.

Architecture

To implement the solution, you will build the following architecture.

Users sign in the AIOps Chatbot using the credentials configured in AWS IAM Identity Center. You will use finding and removing public access from S3 buckets along with finding and closing specific open ports on Amazon EC2 instances as the use cases to demonstrate the capability of this AIOps chatbot using Amazon Q Business custom plugins. However, you can extend the architecture to support other operations use cases through API based integration.

You deploy the required infrastructure using the AWS Serverless Application Model (AWS SAM).

The following is a summary of the functionality of the architecture:

• The UI for the chatbot is built using an Amazon Q Business web experience.
• The user authentication and authorization are handled by AWS IAM Identity Center.
• Relevant actions are identified based on natural language queries from the users using Amazon Q Business custom plugins. Amazon Q Business uses the configured third-party OpenAPI specifications to dynamically determine which API operations to perform to fulfill an end user request.
• The APIs are implemented using Amazon API Gateway and AWS Lambda functions.

Prerequisites

• Create an AWS account if you do not already have one.
• Have access to an AWS account through the AWS Management Console and the AWS Command Line Interface (AWS CLI). The AWS Identity and Access Management (IAM) user that you use must have permissions to make the necessary AWS service calls and manage AWS resources mentioned in this post. While providing permissions to the IAM user, follow the principle of least-privilege.
• Have Git installed.
• Have AWS Serverless Application Model (AWS SAM)
• You must have an Amazon Q Business subscription.
• You must enable AWS IAM Identity Center.
• [Optional] You can pre-create the user in the Identity Center directory that you will be using to sign in to the Amazon Q Business application.

Deploy and run the solution

The resources in this demonstration will be provisioned in the US East (N. Virginia) AWS Region (us-east-1). You walk through the following phases to implement the model customization workflow:

1. Deploy the solution using the AWS SAM template
2. Configure a user for the AIOps Q Business chatbot application
3. Test the AIOps Q Business chatbot application
4. Clean up

Step 1: Deploy the solution using the AWS SAM template

See the GitHub repository for the latest instructions. Run the following steps to deploy the AWS Step Functions workflow using the AWS SAM template.

1. Create a new directory, navigate to that directory in a terminal, and clone the GitHub repository:

git clone https://github.com/aws-samples/ai-ops-with-amazon-q-business.git

2. Change directory to the solution directory:

cd ai-ops-with-amazon-q-business

3. Run the following command to deploy the resources using SAM.

sam deploy -g

4. When prompted, enter the following parameter values:

Stack Name [sam-app]: aiops
AWS Region [us-east-1]: us-east-1
Confirm changes before deploy [y/N]: N

Allow SAM CLI IAM role creation [Y/n]: Y

Disable rollback [y/N]: N

FindS3BucketsWithPublicAccessFunction has no authentication. Is this okay? [y/N]: y

RemovePublicAcessFromS3BucketFunction has no authentication. Is this okay? [y/N]: y

FindEC2WithSpecificOpenPortFunction has no authentication. Is this okay? [y/N]: y

CloseUnwantedPortForEC2Function has no authentication. Is this okay? [y/N]: y

Save arguments to configuration file [Y/n]: Y

SAM configuration file [samconfig.toml]: hit enter

SAM configuration environment [default]: hit enter  

5. Note the outputs from the AWS SAM deployment process. This contains the Amazon Q Business web experience (chatbot) URL. Before you can sign in to the chatbot application, you must set up a user.

Step 2: Configure a user for the AIOps Amazon Q Business chatbot application

Use the following steps to configure a user for the AIOps chatbot application.

1. Open Amazon Q Business from the console and select the AIOps application.

2. Choose Manage access and subscription.

3. Choose Add groups and users.

4. Select either Add and assign new users or Assign existing users and groups depending on if you pre-created the user as mentioned in the prerequisites and choose Next.

5. If you have an existing user that you want to provide access to your AIOps application, search for and select the username and choose Assign.

6. On the review page, select the current subscription and choose Confirm.

Step 3: Test the AIOps Q Business chatbot application

Use the following steps to log into the chatbot and test it. Responses from large language models are non-deterministic. Hence, you may not get the exact same response every time.

1. Take the QBusinessWebExperienceURL from the sam deploy output using the user credential configured in the previous step.
2. After signing in to the AIOps Chatbot, select the kebab menu option (three dots) at the bottom right corner and select the AIOpsCustomPlugin as follows:

3. Enable public access on an Amazon S3 bucket. This is done for testing purposes only, so check your organization policies before performing this test. For this demo we used a bucket named aiops-chatbot-demo.

4. Return to the AIOps Chatbot and enter a question such as: Do I have any S3 bucket with public access? and choose Submit. Provide the bucket prefix to narrow down the search.

5. The AIOps chatbot identifies the buckets that have public access:

6. Ask a follow up question such as: Please block the public access. The chat bot blocks public access. Validate the change from the S3 console.

7. Open a port, such as 1234, for an Amazon EC2 instance using security group inbound rules.

8. Return to the chat bot and enter a question such as: Do I have any EC2 instance with port 1234 open?

9. After the chat bot identifies the EC2 instance with the open port, confirm that you want to close the port.

10. The chat bot closes the open port and confirms.

Clean up

Properly decommissioning provisioned AWS resources is an important best practice to optimize costs and enhance security posture after concluding proofs of concept and demonstrations. To delete the resources deployed to your AWS account through AWS SAM, run the following command:

sam delete

OpenAPI schema definition

After the custom plugin is deployed, Amazon Q Business will process a user’s prompt and use the OpenAPI schema to dynamically determine the appropriate APIs to call to accomplish the user’s goal. Therefore, the OpenAPI schema definition has a big impact on API selection accuracy. Follow the best practices for OpenAPI schema definition for ideal results. This AIOps chatbot demonstrated four operations supported by the following API operations:

• find-s3-bucket-with-public-access – This API finds S3 buckets that have the specified prefix and are configured for public access.
• remove-public-access-from-s3-bucket – This API removes public access from a specific S3 bucket.
• find-ec2-with-specific-open-port – This API finds EC2 instances that have a specified port open for inbound access.
• close-unwanted-port-for-ec2 – This API removes a specified port from a given EC2 instance.

The API operations are implemented using API Gateway and Lambda functions.

Troubleshooting

The following are some troubleshooting steps if you encounter errors while using the AIOps chatbot.

• As Amazon Q Business dynamically determines the appropriate API operations to be invoked, the questions (prompts) must be unambiguous. Be specific rather than asking generic questions. For example: Do I have any EC2 instance with port 1234 open? instead of Do I have any EC2 exposed to internet?
• The APIs are exposed using API Gateway backed by Lambda functions. Check that you can invoke the API operations using Curl or API testing tools.
• Check the Lambda function logs in Amazon CloudWatch for errors. Follow the Lambda debugging steps if needed.

Conclusion

In this post, you learned an end-to-end process for creating an AIOps chatbot using Amazon Q Business custom plugins, demonstrating how users can use natural language processing to interact with AWS resources and streamline cloud operations. By integrating other AWS services with Amazon Q Business, the chatbot can query infrastructure for security and compliance status while automating key actions such as closing open ports or restricting public access to S3 buckets. This solution enhances operational efficiency, reduces manual intervention, and enabled teams to manage complex environments more effectively through intuitive, conversational interfaces. With custom plugins and OpenAPI schemas, users can build a powerful, flexible chatbot solution tailored to their specific operational needs, transforming the way they manage IT operations and respond to business challenges.

Further study

For more information on Amazon Q Business and custom plugins:

• Amazon Q Business
• Custom plugins for Amazon Q Business
• Prerequisites for Amazon Q Business custom plugins
• Defining OpenAPI schemas for custom plugins
• Creating an Amazon Q Business custom plugin
• Using an Amazon Q Business custom plugin
• Best practices for OpenAPI schema definition for custom plugins

About the authors

Upendra V is a Sr. Solutions Architect at Amazon Web Services, specializing in Generative AI and cloud solutions. He helps enterprise customers design and deploy production-ready Generative AI workloads, implement Large Language Models (LLMs) and Agentic AI systems, and optimize cloud deployments. With expertise in cloud adoption and machine learning, he enables organizations to build and scale AI-driven applications efficiently.

Biswanath Mukherjee is a Senior Solutions Architect at Amazon Web Services. He works with large strategic customers of AWS by providing them technical guidance to migrate and modernize their applications on AWS Cloud. With his extensive experience in cloud architecture and migration, he partners with customers to develop innovative solutions that leverage the scalability, reliability, and agility of AWS to meet their business needs. His expertise spans diverse industries and use cases, enabling customers to unlock the full potential of the AWS Cloud.

Read More

This post is co-written with Keith Brazil, Julien Didier, and Bryan Rand from TransPerfect.

TransPerfect, a global leader in language and technology solutions, serves a diverse array of industries. Founded in 1992, TransPerfect has grown into an enterprise with over 10,000 employees in more than 140 cities on six continents. The company offers a broad spectrum of services, including translation, localization, interpretation, multicultural marketing, website globalization, subtitling, voiceovers, and legal support services. TransPerfect also uses cutting-edge technology to offer AI-driven language solutions, such as its proprietary translation management system, GlobalLink.

This post describes how the AWS Customer Channel Technology – Localization Team worked with TransPerfect to integrate Amazon Bedrock into the GlobalLink translation management system, a cloud-based solution designed to help organizations manage their multilingual content and translation workflows. Organizations use TransPerfect’s solution to rapidly create and deploy content at scale in multiple languages using AI.

Amazon Bedrock is a fully managed service that simplifies the deployment and management of generative AI models. It offers access to a variety of foundation models (FMs), enabling developers to build and scale AI applications efficiently. Amazon Bedrock is designed to be highly scalable, secure, and straightforward to integrate with other AWS services, making it suitable for a broad array of use cases, including language translation.

The AWS Customer Channel Technology – Localization Team is a long-standing TransPerfect customer. The team manages the end-to-end localization process of digital content at AWS, including webpages, technical documentation, ebooks, banners, videos, and more. The AWS team handles billions of words in multiple languages across digital assets. Given the growing demand for multilingual content by internationally minded businesses and new local cloud adoption journeys, the AWS team needs to support an ever-increasing load and a wider set of languages. To do so, the team relies on the GlobalLink technology suite to optimize and automate translation processes.

The challenge

The AWS team and TransPerfect created streamlined custom workflows and toolsets that enable the translation and delivery of billions of words each year. Content localization is a multi-step process consisting minimally of asset handoff, asset preprocessing, machine translation, post-editing, quality review cycles, and asset handback. These steps are often manual, costly, and time-consuming. AWS and TransPerfect are continually striving to optimize this workflow to enable the processing of more content at a lower cost and to decrease those assets’ time to market—providing valuable, salient content faster for non-English-speaking customers. Additionally, transcreation of creative content posed a unique challenge, because it traditionally required highly skilled human linguists and was resistant to automation, resulting in higher costs and longer turnaround times. To address these issues, TransPerfect worked with AWS to evaluate generative AI-powered initiatives for transcreation and automatic post-editing within TransPerfect’s GlobalLink architecture.

Security and data safety

Amazon Bedrock helps make sure data is neither shared with FM providers nor used to improve base models. Amazon Bedrock adheres to major compliance standards like ISO and SOC and is also a FedRAMP-authorized service, making it suitable for government contracts. The extensive monitoring and logging capabilities of Amazon Bedrock allow TransPerfect to align with stringent auditability requirements.

Although data safety is a key requirement, there are many other factors to take into account, such as responsible AI. Amazon Bedrock Guardrails enabled TransPerfect to build and customize truthfulness protections for the automatic post-edit offering. Large language models (LLMs) can generate incorrect information due to hallucinations. Amazon Bedrock supports contextual grounding checks to detect and filter hallucinations if the responses are factually incorrect or inconsistent. This is a critical feature for a translation solution that requires perfect accuracy.

Harnessing LLMs for automatic post-editing

To translate at scale, Amazon Translate powered machine translation is used in AWS team workflows. Segments whose translations can’t be recycled from translation memories (databases of previous high-quality human translations) are routed to machine translation workflows. Depending on the language or content, Amazon either uses a machine translation-only workflow where content is translated and published with no human touch, or machine translation post-edit workflows. Post-editing is when a linguist finesses the machine-translated output of a given segment to make sure it correctly conveys the meaning of the original sentence and is in line with AWS style guides and agreed glossaries. Because this process can add days to the translation timeline, automating some or all of the process would have a major impact on cost and turnaround times.

The following diagram illustrates the machine translation workflow.

The workflow consists of the following components:

• TM (translation memory) – The translation memory is a client-specific repository of previously translated and approved content. It’s always applied first and maximizes the reuse of existing translations.
• MT (machine translation) – After existing translations are applied, new content is processed through machine translation using Amazon Translate.
• APE (automated post-edit) – An LLM is employed to edit, improve, and correct machine-translated content.
• HPE (human post-edit) – A subject matter expert linguist revises and perfects the machine-translated content.

The following example follows the path through the preceding workflow for one source segment.

TransPerfect began working with generative AI and LLMs several years ago with the foresight that AI was on track to disrupt the translation industry. As expected, localization workflows have mostly shifted to “expert in the loop”, and are striving toward “no human touch” models. In pursuit of this, TransPerfect chose to use Amazon Bedrock within its GlobalLink Enterprise solution to further automate and optimize these workflows. Amazon Bedrock, by design, provides data ownership and security. This is a critical feature for TransPerfect clients, especially those in sensitive industries such as life sciences or banking.

With Amazon Bedrock and GlobalLink, machine-translated content is now routed through one of the LLMs available in Amazon Bedrock for automatic post-editing. By using style guides, relevant examples of approved translations, and examples of errors to avoid, the LLM is prompted to improve existing machine translations. This post-edited content is either handed off to a linguist for a lighter post-edit (a less difficult task) or is applied in “no human touch workflows” to greatly improve the output. The result is enhanced quality across the board and the ability for post-editors to focus on higher-value edits.

For post-editing, over 95% of all edits suggested by Amazon Bedrock LLMs showed markedly improved translation quality, leading to up to 50% overall cost savings for translations for Transperfect and freeing human linguists for higher-level tasks.

Harnessing LLMs for transcreation

Although machine translation shows great strength in technical, formal, and instructional content, it hasn’t historically performed as well with creative content that leans into nuance, subtlety, humor, descriptiveness, and cultural references. Creative content can sound stiff or unnatural when machine translated. Because of this, TransPerfect has traditionally relied on human linguists to manually transcreate this type of content.

Transcreation is the process of adapting a message from one language to another while maintaining its intent, style, tone, and context. In German, for example, Nike’s “Just do it” tagline is transcreated to “Du tust es nie nur für dich,” which actually means “you never do it just for yourself.”

A successfully transcreated message evokes the same emotions and carries the same implications in the target language as it does in the source language. The AWS team uses transcreation for highly creative marketing assets to maximize their impact in a given industry. However, transcreation historically hasn’t benefitted from the automation solutions used in other types of localization workflows due to the highly customized and creative nature of the process. This means there has been a lot of interest in using generative AI to potentially decrease the costs and time associated with transcreation.

TransPerfect sought to use LLMs to cut down on time and costs typically associated with transcreation. Rather than an all-human or fully automated process, translations are produced through Anthropic’s Claude or Amazon Nova Pro on Amazon Bedrock, with the prompt to create multiple candidate translations with some variations. Within the translation editor, the human linguist chooses the most suitable adapted translation instead of composing it from scratch.

The following screenshot shows an LLM-powered transcreation within the GlobalLink Translate online editor.

Using GlobalLink powered by Amazon Bedrock for transcreation, users are seeing linguist productivity gains of up to 60%.

Conclusion

Thanks to LLM-powered transcreation and post-editing, customers in industries ranging from life sciences to finance to manufacturing have seen cost savings of up to 40% within their translation workflows and up to an 80% reduction in project turnaround times. In addition, the automatic post-edit step added to machine translation-only workflows provides a major quality boost to the no human touch output.

Amazon Bedrock safeguards data by not allowing sharing with FM providers and excluding it from model improvements. Beyond data security, responsible AI is essential. Amazon Bedrock Guardrails allows TransPerfect to customize truthfulness protections for post-editing. To address AI hallucinations, it offers contextual grounding checks to identify and filter inaccuracies—critical for producing precise translations.

Try out LLM-powered transcreation and post-editing with Amazon Bedrock for your own use case, and share your feedback and questions in the comments.

About the authors

Peter Chung is a Senior Solutions Architect at AWS, based in New York. Peter helps software and internet companies across multiple industries scale, modernize, and optimize. Peter is the author of “AWS FinOps Simplified”, and is an active member of the FinOps community.

Franziska Willnow is a Senior Program Manager (Tech) at AWS. A seasoned localization professional, Franziska Willnow brings over 15 years of expertise from various localization roles at Amazon and other companies. Franziska focuses on localization efficiency improvements through automation, machine learning, and AI/LLM. Franziska is passionate about building innovative products to support AWS’ global customers.

Ajit Manuel is a product leader at AWS, based in Seattle. Ajit heads the content technology product practice, which powers the AWS global content supply chain from creation to intelligence with practical enterprise AI. Ajit is passionate about enterprise digital transformation and applied AI product development. He has pioneered solutions that transformed InsurTech, MediaTech, and global MarTech.

Keith Brazil is Senior Vice President of Technology at TransPerfect, with specialization in Translation Management technologies as well as AI/ML data collection and annotation platforms. A native of Dublin, Ireland, Keith has been based in New York city for the last 23 years.

Julien Didier is Vice-President of Technology for translations.com and is responsible for the implementation of AI for both internal workflows and client-facing products. Julien manages a worldwide team of engineers, developers and architects who ensure successful deployments in addition to providing feedback for feature requests.

Bryan Rand is Senior Vice President of Global Solutions at TransPerfect, specializing in enterprise software, AI-driven digital marketing, and content management strategies. With over 20 years of experience leading business units and implementing customer experience innovations, Bryan has played a key role in driving successful global transformations for Fortune 1000 companies. He holds a BA in Economics from the University of Texas.

Read More

The AWS DeepRacer League is the world’s first autonomous racing league, open to anyone. Announced at re:Invent 2018, it puts machine learning in the hands of every developer through the fun and excitement of developing and racing self-driving remote control cars. Through the past 7 years, over 560 thousand developers of all skill levels have competed in the league at thousands of Amazon and customer events globally. While the final championships concluded at re:Invent 2024, that same event played host to a brand new AI competition, ushering in a new era of gamified learning in the age of generative AI.

In December 2024, AWS launched the AWS Large Language Model League (AWS LLM League) during re:Invent 2024. This inaugural event marked a significant milestone in democratizing machine learning, bringing together over 200 enthusiastic attendees from diverse backgrounds to engage in hands-on technical workshops and a competitive foundation model fine-tuning challenge. Using learnings from DeepRacer, the primary objective of the event was to simplify model customization learning while fostering a collaborative community around generative AI innovation through a gamified competition format.

AWS LLM League structure and outcomes

The AWS LLM League was designed to lower the barriers to entry in generative AI model customization by providing an experience where participants, regardless of their prior data science experience, could engage in fine-tuning LLMs. Using Amazon SageMaker JumpStart, attendees were guided through the process of customizing LLMs to address real business challenges adaptable to their domain.

As shown in the preceding figure, the challenge began with a workshop, where participants embarked on a competitive journey to develop highly effective fine-tuned LLMs. Competitors were tasked with customizing Meta’s Llama 3.2 3B base model for a specific domain, applying the tools and techniques they learned. The submitted model would be compared against a bigger 90B reference model with the quality of the responses decided using an LLM-as-a-Judge approach. Participants score a win for each question where the LLM judge deemed the fine-tuned model’s response to be more accurate and comprehensive than that of the larger model.

In the preliminary rounds, participants submitted hundreds of unique fine-tuned models to the competition leaderboard, each striving to outperform the baseline model. These submissions were evaluated based on accuracy, coherence, and domain-specific adaptability. After rigorous assessments, the top five finalists were shortlisted, with the best models achieving win rates above 55% against the large reference models (as shown in the preceding figure). Demonstrating that a smaller model can achieve competitive performance highlights significant benefits in compute efficiency at scale. Using a 3B model instead of a 90B model reduces operational costs, enables faster inference, and makes advanced AI more accessible across various industries and use cases.

The competition culminates in the Grand Finale, where finalists showcase their models in a final round of evaluation to determine the ultimate winner.

The fine-tuning journey

This journey was carefully designed to guide participants through each critical stage of fine-tuning a large language model—from dataset creation to model evaluation—using a suite of no-code AWS tools. Whether they were newcomers or experienced builders, participants gained hands-on experience in customizing a foundation model through a structured, accessible process. Let’s take a closer look at how the challenge unfolded, starting with how participants prepared their datasets.

Stage 1: Preparing the dataset with PartyRock

During the workshop, participants learned how to generate synthetic data using an Amazon PartyRock playground (as shown in the following figure). PartyRock offers access to a variety of top foundation models through Amazon Bedrock at no additional cost. This enabled participants to use a no-code AI generated app for creating synthetic training data that were used for fine-tuning.

Participants began by defining the target domain for their fine-tuning task, such as finance, healthcare, or legal compliance. Using PartyRock’s intuitive interface, they generated instruction-response pairs that mimicked real-world interactions. To enhance dataset quality, they used PartyRock’s ability to refine responses iteratively, making sure that the generated data was both contextually relevant and aligned with the competition’s objectives.

This phase was crucial because the quality of synthetic data directly impacted the model’s ability to outperform a larger baseline model. Some participants further enhanced their datasets by employing external validation methods, such as human-in-the-loop review or reinforcement learning-based filtering.

Stage 2: Fine-tuning with SageMaker JumpStart

After the datasets were prepared, participants moved to SageMaker JumpStart, a fully managed machine learning hub that simplifies the fine-tuning process. Using a pre-trained Meta Llama 3.2 3B model as the base, they customized it with their curated datasets, adjusting hyperparameters (shown in the following figure) such as:

• Epochs: Determining how many times the model iterates over the dataset.
• Learning rate: Controlling how much the model weights adjust with each iteration.
• LoRA parameters: Optimizing efficiency with low-rank adaptation (LoRA) techniques.

One of the key advantages of SageMaker JumpStart is that it provides a no-code UI, shown in the following figure, allowing participants to fine-tune models without needing to write code. This accessibility enabled even those with minimal machine learning experience to engage in model customization effectively.

By using the distributed training capabilities of SageMaker, participants were able to run multiple experiments in parallel, optimizing their models for accuracy and response quality. The iterative fine-tuning process allowed them to explore different configurations to maximize performance.

Stage 3: Evaluation with Sagemaker Clarify

To make sure that their models were not only accurate but also unbiased, participants had the option to use Amazon SageMaker Clarify for evaluation, shown in the following figure.

This phase included:

• Bias detection: Identifying skewed response patterns that might favor specific viewpoints.
• Explainability metrics: Understanding why the model made certain predictions.
• Performance scoring: Comparing model output against ground truth labels.

While not mandatory, the integration of SageMaker Clarify provided an additional layer of assurance for participants who wanted to validate their models further, verifying that their outputs were reliable and performant.

Stage 4: Submission and evaluation using LLM-as-a-Judge from Amazon Bedrock

After fine-tuned models were ready, they were submitted to the competition leaderboard for evaluation using the Amazon Bedrock Evaluations LLM-as-a-Judge approach. This automated evaluation system compares the fine-tuned models against the reference 90B model using predefined benchmarks, as shown in the following figure.

Each response was scored based on:

• Relevance: How well the response addressed the question.
• Depth: The level of detail and insight provided.
• Coherence: Logical flow and consistency of the answer.

Participants’ models earned a score each time their response outperformed the 90B model in a head-to-head comparison. The leaderboard dynamically updated as new submissions were evaluated, fostering a competitive yet collaborative learning environment.

Grand Finale showcase

The Grand Finale of the AWS LLM League was an electrifying showdown, where the top five finalists, handpicked from hundreds of submissions, competed in a high-stakes live event. Among them was Ray, a determined contender whose fine-tuned model had consistently delivered strong results throughout the competition. Each finalist had to prove not just the technical superiority of their fine-tuned models, but also their ability to adapt and refine responses in real-time.

The competition was intense from the outset, with each participant bringing unique strategies to the table. Ray’s ability to tweak prompts dynamically set him apart early on, providing optimal responses to a range of domain-specific questions. The energy in the room was palpable as finalists’ AI-generated answers were judged by a hybrid evaluation system—40% by an LLM, 40% by expert panelists from Meta AI and AWS, and 20% by an enthusiastic live audience against the following rubric:

• Generalization ability: How well the fine-tuned model adapted to previously unseen questions.
• Response quality: Depth, accuracy, and contextual understanding.
• Efficiency: The model’s ability to provide comprehensive answers with minimal latency.

One of the most gripping moments came when contestants encountered the infamous Strawberry Problem, a deceptively simple letter-counting challenge that exposed an inherent weakness in LLMs. Ray’s model delivered the correct answer, but the AI judge misclassified it, sparking a debate among the human judges and audience. This pivotal moment underscored the importance of human-in-the-loop evaluation, highlighting how AI and human judgment must complement each other for fair and accurate assessments.

As the final round concluded, Ray’s model consistently outperformed expectations, securing him the title of AWS LLM League Champion. The Grand Finale was not just a test of AI—it was a showcase of innovation, strategy, and the evolving synergy between artificial intelligence and human ingenuity.

Conclusion and looking ahead

The inaugural AWS LLM League competition successfully demonstrated how large language model fine-tuning can be gamified to drive innovation and engagement. By providing hands-on experience with cutting-edge AWS AI and machine learning (ML) services, the competition not only demystified the fine-tuning process, but also inspired a new wave of AI enthusiasts to experiment and innovate in this space.

As the AWS LLM League moves forward, future iterations will expand on these learnings, incorporating more advanced challenges, larger datasets, and deeper model customization opportunities. Whether you’re a seasoned AI practitioner or a newcomer to machine learning, the AWS LLM League offers an exciting and accessible way to develop real-world AI expertise.

Stay tuned for upcoming AWS LLM League events and get ready to put your fine-tuning skills to the test!

About the authors

Vincent Oh is the Senior Specialist Solutions Architect in AWS for AI & Innovation. He works with public sector customers across ASEAN, owning technical engagements and helping them design scalable cloud solutions across various innovation projects. He created the LLM League in the midst of helping customers harness the power of AI in their use cases through gamified learning. He also serves as an Adjunct Professor in Singapore Management University (SMU), teaching computer science modules under School of Computer & Information Systems (SCIS). Prior to joining Amazon, he worked as Senior Principal Digital Architect at Accenture and Cloud Engineering Practice Lead at UST.

Natasya K. Idries is the Product Marketing Manager for AWS AI/ML Gamified Learning Programs. She is passionate about democratizing AI/ML skills through engaging and hands-on educational initiatives that bridge the gap between advanced technology and practical business implementation. Her expertise in building learning communities and driving digital innovation continues to shape her approach to creating impactful AI education programs. Outside of work, Natasya enjoys traveling, cooking Southeast Asian cuisines and exploring nature trails.

Read More