AWS re:Invent 2025 -Build a GenAI Race Tracker w/ Confluent & AWS Bedrock in less than 60 Min-ANT327

Building a GenAI Race Tracker with Confluent & AWS Bedrock

Current State of Generative AI

• 33% of enterprises will include agentic AI in their applications by 2025
• This will drive 15% automation in day-to-day tasks, freeing up significant time and resources
• 80% of customer service issues will be resolved without human interaction by 2030
• Agentic AI will bring 30% savings in operational costs by 2030

Limitations of Large Language Models (LLMs)

• LLMs lack the necessary context and domain-specific knowledge to be truly useful
• Enterprises struggle to provide the right context to LLMs at the right time
• Prompt engineering and optimization become increasingly complex as the use cases expand

The Role of Event-Driven Architecture

• Enterprises faced similar challenges with monolithic architectures and microservices
• Event-driven architecture, using an event broker like Kafka, solves the problem of loose coupling and scalability
• Confluent provides a full data streaming platform to connect, govern, and process data in real-time

Confluent Intelligence

Confluent Intelligence is a suite of tools that seamlessly integrates LLMs, machine learning, and other AI capabilities into streaming workflows:

Streaming Agents

• Simplifies the development and deployment of AI/ML applications by providing a unified platform for data processing and AI/ML tasks
• Allows agents to directly call out to LLMs like Amazon Bedrock or OpenAI
• Coordinates data processing and AI workflows to improve efficiency and reduce operational complexity

Real-Time Context Engine

• Provides a fully managed context serving layer that continuously builds, updates, and delivers real-time context to AI agents
• Materializes streaming data into in-memory, low-latency views that can be instantly queried
• Ensures data accuracy and prevents drift by automatically re-running impacted data when upstream definitions change

Built-in ML Functions

• Provides easy-to-use SQL functions for time series forecasting, anomaly detection, and other common ML tasks
• Optimized for real-time processing to deliver accurate forecasts and reliable anomaly detection
• Enables operational monitoring, demand forecasting, predictive maintenance, and more

Hands-on Demo: Building a GenAI Race Tracker

The demo showcases how to build a real-time race tracker application using Confluent Cloud, Amazon Bedrock, and MongoDB:

1. Backend Data Generation: The backend generates real-time race data and publishes it to Kafka topics.
2. Streaming Agents: Flink queries the race data, calls Amazon Bedrock for commentary, and publishes the results to a Kafka topic.
3. Anomaly Detection: Flink applies anomaly detection to the race data and publishes the results to a separate Kafka topic.
4. Semantic Search: Flink creates embeddings for user queries, stores them in MongoDB, and performs vector searches to retrieve relevant commentary.

The demo highlights how Confluent Intelligence simplifies the development and deployment of these AI-powered, event-driven applications by:

• Providing a unified platform for data processing and AI/ML tasks
• Automating the management of real-time context and data governance
• Offering built-in ML functions for common use cases

Best Practices for Streaming and GenAI

1. Selective Data Ingestion: Don't send everything to the LLM; be selective to avoid overwhelming the model and incurring unnecessary costs.
2. Understand Service Quotas: Review the service quotas for the LLM models you're using to ensure you don't hit any limits.
3. Leverage Cross-Region Inference: Use AWS's cross-region inference profiles to distribute the load across multiple regions and increase your service limits.

Resources

• Confluent Marketplace: Get started with $400 in credits and discover streaming solutions
• Confluent.io Quick Start: Learn more about Confluent's streaming agent capabilities
• Hands-on Workshops:
  • Confluent and Databricks: "Delta Tables with AI Agents"
  • Confluent: "Streaming Agent Deep Dive"