## Company and Use Case Overview INRIX is a transportation intelligence company that has pioneered the use of GPS data from connected vehicles for over 20 years. The company operates a massive 50 petabyte data lake containing real-time and historical transportation data collected from connected cars, mobile devices, roadway sensors, and event monitoring systems. INRIX serves automotive, enterprise, and public sector use cases, ranging from financial services datasets to digital twins for major cities like Philadelphia and San Francisco. In June 2024, the California Department of Transportation (Caltrans) selected INRIX for a proof of concept to develop a generative AI-powered solution aimed at improving safety for vulnerable road users (VRUs). The core challenge was to harness the combination of Caltrans' asset, crash, and points-of-interest data with INRIX's extensive data lake to anticipate high-risk locations and rapidly generate empirically validated safety measures to mitigate potential crashes. ## Technical Architecture and LLMOps Implementation The solution, built around INRIX's Compass platform announced in November 2023, demonstrates a sophisticated LLMOps architecture leveraging AWS services. The system employs a multi-component approach with distinct phases for countermeasure generation and image visualization. ### Countermeasures Generation Pipeline The countermeasure generation component utilizes Amazon Bedrock Knowledge Bases integrated with Anthropic's Claude models to implement Retrieval Augmented Generation (RAG). This setup enables the system to process natural language queries such as "Where are the top five locations with the highest risk for vulnerable road users?" and "Can you recommend a suite of proven safety countermeasures at each of these locations?" The RAG implementation allows users to probe deeper into roadway characteristics that contribute to risk factors and identify similar locations in the roadway network that meet specific conditions. The system architecture includes Amazon API Gateway and Amazon Elastic Kubernetes Service (Amazon EKS) for managing API requests and responses. This serverless architecture approach demonstrates scalable LLMOps practices, allowing the system to handle varying loads while maintaining performance consistency. The use of Kubernetes for container orchestration suggests a focus on operational reliability and deployment automation. Behind the scenes, the Compass AI system uses foundation models to query the roadway network, identifying and prioritizing locations with systemic risk factors and anomalous safety patterns. The solution provides prioritized recommendations for operational and design solutions based on industry knowledge, demonstrating how LLMs can be effectively applied to domain-specific expertise. ### Image Visualization and Generation The image visualization component represents an innovative application of generative AI in transportation planning. Traditionally, the process of creating conceptual drawings for transportation countermeasures involved multiple specialized teams including transportation engineers, urban planners, landscape architects, CAD specialists, safety analysts, public works departments, and traffic operations teams. This collaborative process typically extended timelines significantly due to multiple rounds of reviews, adjustments, and approvals. INRIX's solution addresses this challenge by implementing Amazon Nova Canvas for image generation and in-painting capabilities. The system uses AWS Lambda for processing requests and Amazon Bedrock with Nova Canvas to provide sophisticated image editing operations. The implementation supports text-to-image generation and image-to-image transformation, enabling rapid iteration of conceptual drawings. ### In-Painting and Few-Shot Learning Implementation The in-painting functionality enables object replacement through two distinct approaches: binary mask-based replacement for precise area targeting, and text prompt-based identification for more flexible object modification. This demonstrates advanced prompt engineering techniques adapted for visual content generation. The system incorporates few-shot learning approaches with reference images and carefully crafted prompts, allowing seamless integration of city-specific requirements into generated outputs. This approach addresses the challenge of maintaining compliance with local standards while accelerating the design process. The few-shot learning implementation suggests sophisticated prompt engineering practices that enable the system to adapt to different municipal requirements without extensive retraining. ## Production Deployment and Operational Considerations The solution follows a two-stage process for visualizing transportation countermeasures in production. Initially, the system employs image generation functionality to create street-view representations corresponding to specific longitude and latitude coordinates where interventions are proposed. Following the initial image creation, the in-painting capability enables precise placement of countermeasures within the generated street view scene. The Amazon Bedrock API facilitates image editing and generation through the Nova Canvas model, with responses containing generated or modified images in base64 format. This approach demonstrates practical considerations for handling large binary data in production LLMOps systems. The base64 encoding approach suggests attention to data transfer efficiency and integration with existing workflows. ## Scalability and Performance Considerations The serverless architecture approach using API Gateway and Lambda demonstrates scalable LLMOps practices. The combination of Amazon EKS for the main application infrastructure and Lambda for specific processing tasks suggests a hybrid approach that balances performance requirements with operational efficiency. The RAG implementation can be extended to incorporate county-specific regulations, standardized design patterns, and contextual requirements. This extensibility demonstrates how LLMOps systems can be designed to accommodate evolving requirements without fundamental architectural changes. ## Safety and Responsible AI Implementation Amazon Nova Canvas incorporates built-in safety measures, including watermarking and content moderation systems, addressing responsible AI implementation concerns. This demonstrates awareness of potential misuse and the importance of maintaining content integrity in production systems. The comprehensive range of image editing operations supported by the system includes basic image generation, object removal, object replacement, creation of image variations, and background modification. This versatility makes the solution suitable for various professional applications requiring sophisticated image editing while maintaining safety standards. ## Operational Impact and Efficiency Gains The integration of generative AI capabilities enables rapid iteration and simultaneous visualization of multiple countermeasures within a single image. Traditional manual visualization processes that previously required extensive time and resources can now be executed efficiently through automated generation and modification. The solution delivers substantial improvements in both time-to-deployment and cost-effectiveness, potentially reducing design cycles from weeks to days. ## Technical Challenges and Considerations While the case study presents significant achievements, certain technical challenges warrant consideration. The reliance on large foundation models introduces latency and cost considerations that must be managed in production environments. The quality and consistency of generated visualizations depend heavily on prompt engineering and the underlying training data of the foundation models. The system's effectiveness in handling edge cases and unusual scenarios may require ongoing refinement and validation. The integration of multiple AWS services introduces complexity in monitoring, debugging, and maintaining system reliability across the entire pipeline. ## Future Extensibility and Integration The modular architecture suggests potential for future enhancements and integrations. The RAG implementation provides a foundation for incorporating additional data sources and domain expertise. The image generation capabilities could be extended to support additional visualization formats and interactive elements. The system's design appears to support integration with existing transportation planning workflows and tools, suggesting practical deployment considerations were addressed during development. The ability to handle city-specific requirements through few-shot learning indicates adaptability to diverse operational environments. This case study demonstrates a sophisticated application of LLMOps principles in addressing complex real-world challenges in transportation planning and safety. The combination of large-scale data processing, natural language understanding, and automated visualization represents a comprehensive approach to leveraging AI in production environments while maintaining focus on practical outcomes and operational efficiency.