This case study presents two distinct but complementary approaches to deploying generative AI in production environments, as demonstrated by Loka and Domo during an AWS GenAI Live episode. The discussion reveals sophisticated implementations of agentic AI systems that go far beyond simple chatbot applications, showcasing how organizations can leverage multiple AI models working in concert to solve complex business problems. **Loka's Approach to Production AI Implementation** Loka, led by Chief Innovation Officer Emily Krueger, has established itself as a leading AWS partner with over 300 customers implementing generative AI solutions, with the majority successfully moving to production. Their approach emphasizes a systematic methodology that begins with careful use case validation before any technical implementation. Rather than accepting every client request at face value, Loka conducts thorough discovery periods to identify truly valuable use cases that will generate positive ROI and can be sustained long-term in production environments. The company's production deployment strategy involves a three-phase approach: discovery and validation, proof of concept development, and production implementation. This methodology addresses one of the most critical challenges in LLMOps - ensuring that AI solutions provide genuine business value rather than serving as technology demonstrations. Loka's emphasis on ROI-positive use cases reflects their understanding that production AI systems must justify their operational costs, particularly given the computational expenses associated with running large language models at scale. **Advanced Drug Discovery Assistant (ADA)** Loka's most sophisticated demonstration involves their drug discovery assistant, which represents a pinnacle of agentic AI implementation. ADA integrates multiple specialized AI models and databases to accelerate pharmaceutical research workflows. The system combines traditional computational biology tools like AlphaFold with newer generative models such as ESM (Evolutionary Scale Modeling) protein family models to create a comprehensive research assistant. The technical architecture of ADA demonstrates several key LLMOps principles. The system maintains connections to multiple external databases including KEGG (pathway databases), STRING DB (protein-protein interactions), and GEO (gene expression databases). This multi-database integration showcases how production AI systems must handle diverse data sources and maintain reliable connections to external APIs. The system's ability to automatically query the appropriate database based on natural language prompts represents sophisticated model routing and orchestration capabilities. ADA's workflow capabilities include target identification through pathway database queries, protein-protein interaction analysis, disease-related expression analysis, and molecular docking simulations. The system can retrieve protein sequences, perform protein folding using state-of-the-art models, and execute blind docking of molecules using diffusion models like DiffDock. This comprehensive functionality demonstrates how agentic systems can chain together multiple AI models and computational tools to complete complex scientific workflows. The production implementation of ADA addresses several critical LLMOps challenges. The system must handle domain-specific scientific data, maintain accuracy across multiple specialized models, and provide researchers with actionable insights. The integration of 3D visualization capabilities for protein structures and molecular interactions shows how production AI systems can provide rich, interactive outputs that enhance user understanding and decision-making. **Model Diversity and Infrastructure Strategy** Both Loka and Domo emphasize the importance of model agnosticism in production environments. Loka's approach involves leveraging different models for different use cases - utilizing Claude for certain types of reasoning tasks, fine-tuning smaller language models for domain-specific applications, and incorporating open-source models where appropriate. This strategy reflects a mature understanding of LLMOps where organizations must balance performance, cost, and capabilities across diverse use cases. The infrastructure strategy heavily leverages AWS services, particularly Amazon Bedrock for model access and Amazon SageMaker for custom model deployment. This approach provides several production advantages including built-in security controls, scalability, and the ability to quickly experiment with new models as they become available. The emphasis on running models within customer VPCs addresses critical security and compliance requirements for enterprise deployments. **Domo's Business Intelligence Agentic Systems** Domo's implementation represents a different but equally sophisticated approach to production AI deployment. Their platform integrates over 1,000 pre-built connectors to various data sources, creating a comprehensive data aggregation and analysis platform. The company's agentic AI implementations focus on automating business intelligence workflows while maintaining human oversight through carefully designed human-in-the-loop systems. The call center optimization use case demonstrates real-time agentic AI deployment. When a customer calls, the system automatically triggers API calls to Domo's agentic workflow, which then queries multiple data sources including customer databases, flight information systems, and rebooking platforms. The agent synthesizes this information to provide call center representatives with comprehensive customer context, suggested scripts, and resolution options before the call even begins. This implementation showcases how production AI systems can integrate with existing business processes and provide just-in-time intelligence to human operators. The financial analysis use case represents batch processing capabilities where agentic systems analyze profit and loss statements on a scheduled basis. The system performs root cause analysis, identifies critical concerns, and generates prioritized action items for business stakeholders. This demonstrates how production AI can handle complex financial data and provide executive-level insights without requiring manual analysis. **Production Architecture and Workflow Design** Domo's workflow architecture provides insights into production agentic system design. Their agent tasks include four key components: input specifications, tool definitions, role and task descriptions, and output specifications. This structured approach ensures that agents have clear objectives and appropriate tools to accomplish their tasks. The tool definitions include SQL query capabilities, API integrations, and data transformation functions, demonstrating how production agents require diverse capabilities to handle real-world business scenarios. The human-in-the-loop implementation represents a crucial aspect of production AI deployment. Rather than fully automating processes, both companies emphasize the importance of human oversight and validation. Domo's task queue system allows human operators to review agent outputs, make corrections, and provide additional context before final actions are taken. This approach balances automation benefits with the need for human judgment and accountability. **Data Management and Security Considerations** Both companies address critical data management challenges in production AI deployment. Loka emphasizes the importance of data readiness, including data accessibility, cleanliness, and security compliance. They particularly highlight the need for proper handling of PII, HIPAA compliance in healthcare applications, and regulatory requirements that are increasingly important as AI governance laws emerge. The security architecture leverages AWS's security model, with models running entirely within customer VPCs and data never leaving the security boundary. This approach addresses one of the primary concerns organizations have about production AI deployment - ensuring that sensitive data remains protected while still enabling AI capabilities. **Evaluation and Monitoring Frameworks** The case study reveals sophisticated approaches to evaluation and monitoring in production environments. Loka emphasizes the importance of baseline performance metrics and ground truth data for model evaluation. They advocate for automated evaluation frameworks that can assess model performance without requiring constant human review, while still maintaining human-in-the-loop capabilities for critical decisions. Domo's approach includes performance monitoring that tracks time savings and business impact. Their call center implementation reportedly achieves up to 50% time savings per call, demonstrating measurable business value. This focus on quantifiable outcomes reflects mature LLMOps practices where organizations track both technical performance metrics and business impact measures. **Migration and Multi-Model Management** The discussion reveals increasing demand for AI migration services, particularly as organizations seek to avoid vendor lock-in and optimize their model portfolios. Loka's experience with migration projects demonstrates how production AI systems must be designed for flexibility and portability. Their approach involves creating infrastructure that can support multiple models simultaneously, enabling organizations to route different types of queries to the most appropriate model based on complexity, cost, and performance requirements. The model routing capabilities described by both companies represent sophisticated production features where systems automatically determine which model or combination of models should handle specific requests. This optimization reduces costs by avoiding unnecessary use of expensive models for simple queries while ensuring complex requests receive appropriate processing power. **Scaling and Operational Considerations** Both companies address the operational challenges of scaling AI systems in production. Domo's connector architecture demonstrates how production systems must handle diverse data integration requirements while maintaining performance and reliability. Their ability to support real-time data updates (as frequently as every 15 minutes) shows how production AI systems must handle dynamic data environments. The scaling strategies involve careful consideration of computational costs, model performance, and user experience. Both companies emphasize the importance of building systems that can handle production workloads while maintaining cost-effectiveness. This includes strategies for model caching, request batching, and intelligent routing to optimize resource utilization. **Future Directions and Maturity Curve** The case study reveals a clear maturity curve in AI adoption, with organizations moving from simple chatbot implementations to sophisticated multi-agent systems. Both companies predict that future developments will involve increased automation with reduced human involvement, but only after organizations build confidence in AI capabilities through initial human-in-the-loop implementations. The discussion of agentic systems represents the current frontier of production AI deployment, where multiple AI agents work together to complete complex workflows. This approach enables organizations to tackle sophisticated business problems that require coordination across multiple systems and data sources, representing a significant advancement over single-model implementations. The emphasis on deterministic workflows coupled with AI capabilities suggests that successful production implementations will continue to balance AI flexibility with process reliability, ensuring that organizations can achieve the benefits of AI automation while maintaining the predictability and control required for business-critical operations.