## Overview Navismart AI has built a production AI system designed to automate immigration processes, a domain characterized by high complexity, sequential workflows, strict regulatory requirements, and high-stakes decision making. The presentation was delivered by Amulo Washington, a product developer at Navismart AI based in Nairobi, Kenya, as part of what appears to be a technical conference or meetup. The team consists of approximately six people split between the US (including the founder) and Kenya (two developers), making this a distributed development effort with a remarkably small team tackling an ambitious problem space. The company chose immigration automation as their initial use case specifically because it represents one of the most complex regulatory environments globally. They began with US immigration processes deliberately, reasoning that if they could handle the strictest and most complex regulatory framework, they would be better positioned to expand to other countries. This strategic decision reflects a thoughtful approach to building robust, production-ready AI systems that must handle real-world complexity and regulatory constraints. ## Problem Space and Challenges The immigration automation domain presents several interconnected challenges that make it particularly demanding for AI systems. First, the workflows are inherently sequential and conditional, requiring multiple stages from document verification through applicant interviews to legal compliance checks. These processes traditionally depended heavily on human expertise, particularly from immigration lawyers and legal advisors. The challenge isn't simply automating individual tasks but orchestrating complex multi-step processes where each stage may have conditional logic and dependencies on previous steps. Second, regulatory compliance varies significantly across different jurisdictions. Different countries enforce different rules at their borders and within their immigration platforms. This creates a need for strict privacy requirements to protect customer and user data, as well as decision transparency mandated by various governance frameworks for data processing. The regulatory landscape isn't static either—it changes over time and varies by region, requiring systems that can adapt to evolving requirements. Third, the system needs reliable performance with predictable operations under variable conditions. This includes seamless integration between AI capabilities and human oversight, particularly for complex cases that AI agents cannot fully handle autonomously. The high-stakes nature of immigration decisions means that errors can have serious consequences for individuals, demanding both accuracy and transparency in how decisions are made. ## Multi-Agent Architecture Design Navismart AI implemented a modular microservices architecture with multiple specialized agents, each handling different operations. This design decision reflects a deliberate architectural choice to break down the complex immigration workflow into manageable, independently deployable components. Each agent specializes in a specific task: one agent handles form filling automation, another performs verification processes, another reviews documents for accuracy and completeness, and so forth. The agents communicate through REST APIs, with message queues enabling seamless inter-agent communication and state synchronization. The architecture follows a hierarchical supervision pattern with a single supervisor agent overseeing multiple worker agents (labeled as worker agent A, worker agent B, worker agent C, etc.). Worker agents have distinct responsibilities—for example, one worker agent's role is specifically to bring in data from external sources and provide reviews and feedback on that data. This architectural pattern represents a peer-to-peer communication model enabled by RESTful APIs for direct agent interaction, which enables flexible collaboration and adaptive workflow responses. The team chose REST APIs specifically because they found them "much more faster" and less bulky during the coding process, with good performance characteristics. When a document validation agent rejects input, for instance, it triggers a revalidation workflow coordinated with a notification agent, creating resilient systems that leverage the specialized strengths of individual agents while maintaining coordination. ## Framework Development and Decision Rationale Interestingly, Navismart AI built their own framework rather than adopting an existing multi-agent orchestration framework. This decision stemmed from their specific requirements around scalability and fault isolation. As explained during the presentation, using a single agent framework to handle all processes would create a significant risk: if a breakdown occurred in a single agent handling everything, it would interfere with other concurrent processes and require bringing down the entire system to debug and fix the issue. By breaking functionality into specialized agents, each handling its own function, the team achieved easier debugging and better observability during the scaling process. They could log and monitor how each agent was performing, identify which agents were handling their processes well, and pinpoint areas needing improvement. This modular approach also provides technology diversity, allowing different frameworks per agent while maintaining fault isolation to capture errors and log them effectively during agent interactions. However, the team didn't build everything from scratch. As they scaled to different countries and their requirements evolved, they integrated with LangGraph for monitoring and evaluation purposes. This integration came as LangGraph was also growing its platform capabilities, and it provided the team with better tools for evaluating LLM performance across different platforms and understanding how their agents were handling processes as they scaled to different regions and countries. This hybrid approach—custom orchestration framework combined with established tools for specific capabilities—represents a pragmatic balance between control and leveraging existing ecosystem solutions. ## Stateful Session Management and Context Preservation A critical LLMOps capability in the Navismart system is stateful session management with collective context preservation across conversation history and decision trees. During chat interactions, the system saves conversation history, enabling agents to remember where a user left off and pick up from that point in subsequent interactions. This persistent memory is essential for immigration processes that may span multiple sessions and require users to provide information gradually over time. The system implements what they describe as "multi-turn dialogue preservation with intent parsing and conversation statement management." This suggests sophisticated natural language understanding capabilities that go beyond simple question-answering to maintain complex conversational state across extended interactions. For processes that may involve multiple forms, documents, and decision points over days or weeks, this context preservation is fundamental to providing a coherent user experience. The system also implements automated feedback loops that retrain models using user interaction data. This is particularly important for features like opportunity searches, where the agent picks up user profile information and uses it to find the best opportunities based on whatever the user has requested. This suggests a degree of personalization and continuous learning built into the production system, though the presentation doesn't provide detailed specifics about the retraining pipeline or how frequently models are updated. ## Deployment Infrastructure and Scaling For deployment and scaling, Navismart AI uses Kubernetes as their orchestration platform. This choice was driven by specific requirements for scalability, isolation, and dynamic instantiation of agents. The containerized services approach allows Kubernetes to orchestrate with autoscaling capabilities, responding to demand fluctuations as the number of users on the platform grows. This is crucial for a service that may experience variable load patterns depending on application deadlines, policy changes, or other factors that drive immigration activity. The team implements CI/CD pipelines with blue-green deployment strategies, incorporating automated testing and minimal downtime release cycles. This production-grade deployment approach ensures that they can identify where to debug issues and where to pick up errors when they occur. The comprehensive monitoring system is also automated and sends alerts to the team, with the presenter specifically thanking someone named "Nate" for bringing automation to the process and making it simpler to get clear real-time information during tracking processes. The independent deployment capability for different agents is a key architectural feature. Each agent can be deployed independently with its own capabilities, providing technology diversity that allows different frameworks per agent while maintaining fault isolation. This microservices approach provides operational flexibility but also introduces complexity in managing service dependencies, version compatibility, and inter-service communication at scale. ## Document Processing and OCR Integration For document processing, a critical component of immigration workflows, Navismart AI integrates Google OCR AI. According to the presenter, Google OCR is "still the best one currently for now" for viewing documents and reading document processes, including on images. This suggests they evaluated different OCR solutions and found Google's offering to provide the best accuracy for their use case, which likely includes a wide variety of document types, formats, and quality levels submitted by applicants from different countries. The document review agent checks documents for accuracy and identifies missing elements in the process. This capability is essential for immigration workflows where incomplete or incorrect documentation can lead to application rejections or delays. The system appears to provide automated feedback to users about documentation issues, though the presentation doesn't detail exactly how this feedback is communicated or what level of specificity is provided about required corrections. ## Country-Specific Regulatory Handling Managing different countries' regulatory frameworks represents one of the most complex aspects of the system. As mentioned, the team deliberately started with the United States because it has some of the strictest immigration rules globally. The reasoning was strategic: if they could handle US complexity first, expanding to other countries would become easier. This approach of tackling the hardest case first as a foundation for simpler cases represents sound engineering judgment, though it also likely meant a longer initial development cycle. The team worked closely with a legal advisor specializing in immigration processes who guided them through every process before implementing features for different countries. This human expertise integration during development—not just during production use—ensured that the automated system correctly captured regulatory requirements and compliance needs. Each service has its own file handling the specific services requested, and when scaling to different countries, they improved the system by evaluating LLM performance on different platforms, which is where their LangGraph integration became valuable. The presentation doesn't provide extensive detail about how country-specific logic is actually implemented in the codebase—whether through configuration files, separate deployment instances, or some other mechanism. This is an area where more technical detail would be valuable for understanding the LLMOps practices around managing regulatory variation at scale. ## Human-in-the-Loop Integration Despite the automation focus, Navismart AI emphasizes human-in-the-loop capabilities for critical decision checkpoints. The system is designed to maintain human oversight for high-stakes operations, with advisory processes and legal reviews integrated where AI agents might miss important considerations. Individual agents operate autonomously while coordinated through structured protocols, but the architecture explicitly includes pathways for human intervention. The system implements "seamless escalation" where complex queries automatically escalate to human agents without disruption. Advanced conversation state management handles these interruptions gracefully while maintaining context, creating what the team describes as natural conversational experiences through persistent memory and real-time adaptation to user needs. This suggests sophisticated handoff protocols between AI and human agents, though again, the technical implementation details aren't fully specified in the presentation. The human oversight dimension extends beyond just handling edge cases. It represents a fundamental architectural principle that recognizes the limitations of current AI systems in high-stakes scenarios with complex regulatory requirements. The presenter explicitly notes that "we can't be sure of everything being on a clean text," acknowledging the messiness of real-world data and the need for human judgment in certain situations. ## Voice Integration and Multilingual Support The team is working on real-time voice agent capabilities featuring speech-to-speech interaction, though it appears this is still in development rather than fully deployed. The vision is to automate the entire process for users who aren't able to interact directly with text-based interfaces. Currently, the system supports speech-to-text and text-to-speech engines optimized for sub-100 millisecond response times, indicating a focus on low-latency processing essential for natural voice interactions. The platform implements multilingual support with real-time translation capabilities, able to translate back to English or whatever language the user prefers. This is crucial for a global immigration platform where users may speak many different languages. The natural language understanding capabilities include multi-turn dialogue preservation with intent parsing and conversation state management, suggesting relatively sophisticated NLU capabilities beyond simple command recognition. The voice integration work represents an interesting evolution in the platform's accessibility, potentially opening it to users with different literacy levels or physical abilities. However, voice interfaces also introduce additional complexity for LLMOps, including managing speech recognition errors, handling various accents and speaking styles, and maintaining conversational coherence across modalities. ## Security, Privacy, and Compliance Given the sensitive nature of immigration data, Navismart AI implements comprehensive security measures. The system features end-to-end encryption with a zero-trust architecture and data anonymization. Auditable trails with detailed logging support regulatory compliance requirements, which vary across the different jurisdictions where the platform operates. The presenter emphasized that security forms "the base of everything now in the field of AI because everyone is now looking into where is my data being used to, where is my data going to, what are they using my data for." The system implements what's described as "interpretable AI decisions with reasoning chains," suggesting some form of explainability built into the decision-making processes. This transparency is essential both for regulatory compliance and for building user trust in high-stakes immigration decisions. The system also uses differential privacy and secure multi-party computation to provide additional safeguards, though the presentation doesn't detail exactly how these privacy-preserving techniques are implemented or what tradeoffs they involve. Regular security audits are part of the operational practice, ensuring ongoing compliance and trust-building. This suggests a mature approach to security operations, though the frequency of audits and whether they're conducted internally or by third parties isn't specified. The human oversight integration also serves a security and compliance function, with human review at critical decision-making checkpoints helping ensure that automated decisions meet regulatory standards. ## Operational Practices and Lessons Learned The team identified several key lessons from building and operating the system in production. First, they learned the importance of balancing automation with human oversight through robust fallback strategies. This echoes a common theme in production AI systems: pure automation often fails in edge cases, and graceful degradation to human handling is essential for reliability. Second, they found that disciplined change management was essential for version compatibility as they evolved from early versions to production-ready systems and then to a full platform usable across different regions. This challenge is common in evolving AI systems where models, prompts, and orchestration logic may all change over time while maintaining backward compatibility with existing workflows. Third, continuous monitoring and rapid incident response proved crucial for reducing disruptions and improving user satisfaction. The automated monitoring system with real-time alerts enabled the team to identify and address issues quickly, which is particularly important for a small team supporting a complex production system. The ability to track how each agent is performing independently provides granular observability that aids in both debugging and optimization. ## Evaluation and Monitoring The integration with LangGraph for evaluation represents an important dimension of the team's LLMOps practice. As they scaled to different countries, they needed better tools to evaluate LLM performance across different platforms and understand how their agents were handling processes. LangGraph provided monitoring capabilities that helped them assess agent performance and make data-driven decisions about system improvements. However, the presentation doesn't provide extensive detail about specific evaluation metrics, whether they use human evaluation, automated evaluation against test sets, or some combination. For an immigration system where accuracy and compliance are critical, one would expect rigorous evaluation protocols, but these aren't fully described. Similarly, the automated feedback loops for retraining models using user interaction data suggest ongoing learning, but the specifics of data collection, annotation, retraining frequency, and deployment of updated models aren't detailed. ## Future Roadmap and Expansion Plans Looking forward, Navismart AI has several directions for development. They plan to standardize agent protocols for deeper ecosystem integration, suggesting a move toward more interoperability with external systems and services. They're working on advanced planning techniques for complex decision futures, which may involve more sophisticated reasoning capabilities or multi-step planning. The team is enhancing multilingual and sentiment-aware interfaces for global accessibility, expanding beyond simple translation to understanding emotional context in user interactions. This could help the system better identify users who are frustrated, confused, or in distress and respond appropriately or escalate to human agents. Beyond immigration, the team is exploring expansion into education processes, particularly helping people who want to move to other countries to enhance their education. They're also looking at employment use cases, potentially helping companies manage employee relocations to different countries where they have operations. These adjacent domains share some similarities with immigration (complex documentation, regulatory requirements, multi-step processes) while also introducing new domain-specific challenges. ## Critical Assessment and Open Questions While the presentation provides a good overview of the system architecture and approach, several areas warrant critical examination and would benefit from more detailed disclosure. First, the actual performance metrics of the system in production aren't provided. What percentage of cases can be handled fully automatically versus requiring human intervention? What is the error rate in document processing or compliance checking? How satisfied are users with the automated experience compared to traditional processes? Second, the evaluation and testing methodology remains somewhat opaque. For a high-stakes domain like immigration, rigorous testing and validation would seem essential, but the presentation doesn't describe in detail how they ensure system accuracy and reliability before deploying changes. The mention of automated testing in the CI/CD pipeline is encouraging but doesn't provide specifics about test coverage, types of tests, or pass criteria. Third, the decision to build their own framework rather than using existing multi-agent orchestration frameworks like CrewAI, AutoGen, or LangGraph's native orchestration is interesting but raises questions. While the stated reasons around fault isolation and scalability make sense, this decision also means maintaining custom infrastructure code that could potentially benefit from community development and standardization. The tradeoff between control and maintenance burden isn't fully explored. Fourth, the handling of model updates and versioning across multiple specialized agents could introduce significant complexity. When you have multiple agents, each potentially using different models or versions, ensuring consistent behavior and managing coordinated updates becomes challenging. The presentation mentions version compatibility as a key lesson but doesn't detail how they actually manage this in practice. Fifth, the security and privacy claims, while reassuring, would benefit from more technical specificity. What encryption standards are used? How is the zero-trust architecture implemented? What exactly do they mean by differential privacy in this context, and what privacy guarantees can users actually expect? These questions are particularly important given that immigration data is highly sensitive and potentially valuable to various actors. Finally, the small team size (six people total) building such an ambitious, complex system is both impressive and potentially concerning. While small teams can move quickly and maintain coherent architectural vision, they may also face challenges with sustainability, knowledge concentration, and ability to handle operational incidents. The distributed nature of the team across Kenya and the US adds additional coordination challenges. ## Broader Context and Transferability The presenter was asked whether lessons from this immigration automation system could extend to other domains. The response focused on domain expansion (education, employment) but the question touches on something broader: what aspects of this multi-agent architecture and LLMOps approach are transferable versus domain-specific? The architectural patterns—specialized agents communicating via REST APIs, hierarchical supervision, containerized deployment with Kubernetes, CI/CD with blue-green deployments, human-in-the-loop for high-stakes decisions—are all broadly applicable to many domains involving complex workflows, regulatory requirements, and document processing. E-commerce, financial services, healthcare, and many other industries could potentially benefit from similar approaches. However, the deep domain expertise required for immigration (working closely with legal advisors, understanding country-specific regulations, handling sensitive personal data) suggests that successfully deploying such systems requires substantial domain-specific investment beyond the technical infrastructure. The transferability may be more at the architectural pattern level than at the system level. ## Conclusion Navismart AI's immigration automation platform represents an ambitious production deployment of multi-agent AI systems in a complex, high-stakes domain. The team has made thoughtful architectural decisions around modularity, fault isolation, and human-in-the-loop integration that reflect awareness of the challenges of production AI systems. Their choice to start with the most complex regulatory environment (US immigration) and use that as a foundation for expansion shows strategic thinking about building robust, scalable systems. However, several aspects of the system would benefit from more detailed disclosure, particularly around performance metrics, evaluation methodologies, and the specifics of how they handle model versioning and updates across multiple agents. The small team size and custom framework decision introduce both agility and potential sustainability questions. As the system expands to more countries and potentially other domains, maintaining quality and consistency while scaling will likely present ongoing challenges. The case study illustrates both the potential and the complexity of bringing LLMs and multi-agent systems to production in regulated, high-stakes environments. It demonstrates that even relatively small teams can build sophisticated systems when they make smart architectural choices and leverage existing tools (Kubernetes, Google OCR, LangGraph) strategically while building custom components where needed. At the same time, it raises important questions about evaluation, testing, security, and operational sustainability that are relevant to anyone building production AI systems in complex domains.