Agents With Real Stakes: Deploying AI for Immigration at Scale // Amolo Washington

Amolo Washington [00:00:06]: Thank you. So for quick introduction, my name is Amola Washington, product developer at NaviSMAT AI that is. So I'll be taking you through deploying AI agents for scalable immigration automation as a big topic now being the immigration automation processes. So let's dive in. So this is the technical journey that I'll be taking you through. Down from immigration automation challenges back up to the last one. That is lessons that we'll be learning about the future outlook of the immigration process, The technical architecture of our product and the implementation process. So however much immigration process might be able to be automated in a simple process, or there might be challenges that always comes in, that is whereby we have the complex sequential workflows, have the conditional logic and the sequential processing where we need to involve the human oversight that is in between.

Amolo Washington [00:01:27]: However, we'll be leaving the processes to the AI agents. We can't be sure of everything being on a clean text. Then we'll have another challenge that is the regulatory compliance which usually changes based on different regions. So different countries usually enforce different rules on their immigration platforms or the borders when you have to access or even within their platforms. So we'll need a strict privacy requirements in order to protect our customer data, even the user's data. And also the decision transparencies which are mandated govern all the data process handling. Again, the next challenge that you always go through, that is the reliable performance. So in order to have a predictable operation and a variable conditions with seamless human AI integration, that is a human and an AI in between, whereby in complex tasks that the AI agent can't be able to handle, we bring in the human intervention that that is where we need the lawyers and the legal advisers during the processes.

Amolo Washington [00:02:49]: So immigration might involve multiple sequential stages, that is from document verification, applicant interviews, legal compliance checks. That traditionally depended on the human expertise with the agents. Now these processes have been automated and made simpler, which requires a careful system design and the architecture, the regulatory adherence and also the operational resilience. So that's where NAVISMAT AI comes in. So our platform is built in modular microservice design where we have multiple agents handling different operations. Operations. Each agent is specialized on a given task. So we have the specialized agent module where the one agent does a given tasks.

Amolo Washington [00:03:54]: Even when we have the automation process, there is an agent that is being that is able to handle the process. For form fielding automation, we'll have to. We'll have an agent to do the verification process. We'll have an agent to review the documents, whether they are accurate or not or you've missed something in the process. So that's where we bring in the agent modules, the specialized agent module. Each agent having a special task assigned to it. Here the agents Communicate through REST APIs where the message queues enable seamless interagent communication and state synchronization. So we'll have a single supervisor agent with multiple agents under it.

Amolo Washington [00:04:51]: We have worker agent A, worker agent B, worker agent C. Then we have the external data sources whereby the worker agent D, B. That's its role to bring in sources from different platforms and also give you a review of whatever it has explored and the feedback that is coming in. So we'll have. We also do the stateful session management I.e. collective context preservation across conversation history and decision trees. Where during the chat interactions your chats are being saved in a history whereby the agent can be able to remember where you left from and pick it up from there. Okay.

Amolo Washington [00:05:42]: Independent deployment. We do have independent deployment of different agents and their capabilities whereby the technology diversity which allow different frameworks per agent, while fault isolation also is big prevented to be able to capture any error or logging of errors during the processes of interactions of the agents. So this is how our multi agent collaboration patterns occur. We have the hierarchical supervision where we have a single agent which supervises the specialized agent agents processes. This one is the margin manager now of the agent which is to oversee the worker agents for coordinated task and execution. We have the peer to peer communication whereby we use the restful APIs for communication between different agents. So the direct agent interaction enables the flexible collaboration and adaptive workflow responses. Again, as we said, we must bring in the human in between for now the advisory processes or the legal processes that might be missed during by the agent.

Amolo Washington [00:07:09]: So that's where bringing the human in the loop. We need the critical decision checkpoints to maintain human oversight. For high stakes and of operations. The individual agents operate autonomous autonomously while coordinated through structured protocols. When a document validation agent rejects input, it triggers revalidation workflow coordinated with notification agent, creating resilient systems that leverage specialized strengths. So each agent is dependent on the other. Like if one has missed something during the process. Now it pushes a notification back to the backend.

Amolo Washington [00:07:55]: Now for the handling of the issue. For the model of the deployment and scaling, we use kubernetes because of its scalability and instantiation of the agents also. So we have the containerized services the kubernetes orchestrate with auto scaling capability, responding to demand and fluctuations due to the growing number of users within the platform. So that's the architecture of how the agents have been put within different microservices. We have the CICD pipelines where we have the blue green deployments with automated testing and minimal downtime release cycles. So in case of errors, you know where you debug from, you know where to pick up the error from. There is a comprehensive monitoring which is automated also and also sends an alert to an end thanks to anything. That brought the automation process much more simpler and brought in the process of now getting clear real time information during the tracking processes.

Amolo Washington [00:09:32]: So we have the feedback, the automated feedback loops which retrain models using user interaction data. This one mostly comes in with what you usually ask the agent even during the opportunity searches. It picks up your profile the way you've put it and and keeps uses your information now to find for you the best opportunities based on whatever you've requested from it. Our next feature that we are very much looking forward to bring in was a real time voice agent whereby we have text speech to speech agent. So automating all the process, this one was to bring in the users who are not able to interact directly with the interface that is typing the text, but instead they can just communicate their info. Then the agent carries the work on the next level. So the voice and chat integration architecture we have the low latency processing where the speech to text, currently it's still speech to text and text to speech engines which optimizes the sub of 100 milliseconds response times. We have the multilingual support whereby it supports different languages globally.

Amolo Washington [00:11:13]: So the real time translation can either translate back to English or the language that you feel like you are okay with. The natural language understanding and context management where we have the multi turn dialogue preservation with intent passing and conversation statement management. Then there is seamless escalation where complex queries automatically escalate to human agents without disruption. So advanced conversation state management handles interruptions gracefully while maintaining context. This creates natural conversational experiences through persistent memory and real time adaption to user needs. As much has been said, we also look into the security because this one 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 you using my data for? Yeah, so we have the end to end encryption whereby we use the zero trust architecture with data anonymization. There is an auditable trails, detailed logging for regulatory compliance, human oversight where we bring them in for the critical decision making process for the checkpoints and human review There is a transparency where interpretable AI decisions with reasoning chains also. So there is differential privacy and secure multi party computation which provides additional safeguards while regular security audits ensure ongoing compliance and trust building.

Amolo Washington [00:13:16]: So having that as much their key lessons during the production process and the future outlook that you are looking forward to have within the platform. The key lessons during the building process of Navismart AI was the balance between automation with human oversight through the robust fallback strategies, the disciplined change management which was essential for the version compatibility from better version now to ready to production use and now a full platform that could be used across every place. The future roadmap that we have is to standardize the agent protocols for deep ecosystem integration to do the advanced planning technique for complex decision in future. There is multilingual and sentiment aware where the interfaces for global anti accessibility. The continuous monitoring and rapid incident response has reduced disruptions and improved user satisfaction while advancing toward more autonomous intelligent systems. I think that's all on my end. Awesome welcoming.

Adam Becker [00:14:51]: Thank you very much. Let's see, do we have folks in the. In the chat. Just submit some questions if you have any. I'm a little. Thank you very much for the presentation. You guys seem to be doing an enormous amount of things. Yeah, this is a massive, massive system.

Adam Becker [00:15:14]: How many people are working on a project like this?

Amolo Washington [00:15:18]: Yeah, just to build us.

Adam Becker [00:15:20]: Oh my God. That's absolutely wild. Yeah, so let me just start then you showed us a diagram. Maybe you can go back to that diagram that I thought was clipped. You know what, you could even share your screen. That might actually be better because I think that you're right now you're using the slides internally. If you want to share your screen and just show that. Yep, exactly.

Adam Becker [00:15:55]: I'll turn this one off. And then as soon as you have your screen up, we can just feel comfortable that we're seeing the right thing.

Amolo Washington [00:16:02]: Okay, cool. Yes, yes, yes, exactly this one.

Adam Becker [00:16:06]: All right, so we see the architecture here. Did you guys build your own framework? What did you do for a framework? Did you pick one and how did you go through the process of selection?

Amolo Washington [00:16:18]: We did our own framework at first because we needed to get a better understanding on what is missing on the immigration processes and, and at least to get the data first from the users on their pain points and everything. So in order for us to come up with now the idea for Reach now to be a production ready, that is the to build in an agent that could be able to automate all the process, then we decided on which agents to carry on which function. So as you could see, the Google OCR AI is still the best one currently for now. The view on the document and the reading of the document processes, even on the images for the deep search.

Adam Becker [00:17:16]: But I guess Amalo, I mean like even just the orchestration of them and how all of them work together. Did you. You created your own framework for how to stitch them together and have them communicate and be orchestrated and supervised or. Yeah. Interesting. And you felt like you needed to because of your relatively specific type of complexity that you're dealing with. Like what was insufficient about let's say like other frameworks.

Amolo Washington [00:17:51]: It's more during the scaling up process because you might incur a problem during the process on a given end, let's say a break in on a single agent that you have. You are using a single agent for the whole framework now to do everything in case of a break on it. Meaning it's going to interfere with other concurrent processes. Yeah. So you have to handle that one in order to bring back everything. But when we broke them into small each agent handling each specialized function, it was easier now to debug. And also during the scalability process you could be able to log and see, monitor how each of every agent is working on and which one is handling its process as well and what is supposed to be improved during the process also.

Adam Becker [00:18:54]: I see.

Amolo Washington [00:18:55]: Okay.

Adam Becker [00:18:55]: I'm trying to make it a little bit larger if people want to be able to see this and we're going to be right here underneath it. Okay. How did you end up managing different countries? So each country has their own very specific regulatory framework and requirements environments. How did you do that? We have a question here from Andreas.

Amolo Washington [00:19:17]: Okay. At first we started with the US being United States having the strict rules more than any other country based on immigration process. So we wanted to do with them first because our system was at first being working on just immigration process for the American. Those who are planning to move to America based on different things. Yeah. So we were done with us then when we moved to other countries now it became easier and having a legal advisor in for the immigration process, we were able to also get a better view of everything. So he was guiding us through every process that you are taking before doing anything for the different countries. Yeah.

Adam Becker [00:20:11]: Okay. So you started with the US because they have a very complex immigration process and maybe they have it, maybe it's complex so that you can have a good, you know, use case and then if you could figure out the US you can figure out the rest. Yes, that's the Idea So you started there, but then how do you actually implement the difference is each one deployed in its own cluster. You have different files to manage very specific workflows that are unique. Like is there any complexity at the level of the code and the architecture and if so, how is it manifesting?

Amolo Washington [00:20:47]: Each service has its own. Each service or its own file that's handling each of every services that's being requested from. So with us currently when we were scaling to different countries now we improved our system by trying to do with evaluation of the LLMs on different platforms. That is we were trying to. Okay, that's when now the Landgraph was also trying to grow its platform. Yeah. So we integrated our system now again into the Landgraph platform in order to do the monitoring and see how our agent is handling processes and how we can scale into different regions and countries also.

Adam Becker [00:21:40]: So yeah, I have a question about. So this is high stakes scenario. It's a fairly complex set of processes and regulations that you have to jump through. Are there any lessons that can extend beyond immigration, let's say E commerce, any other domain?

Amolo Washington [00:22:03]: Yeah, yeah, because we are trying to look into the education process platform too because that's where most people trying to move from to other countries for and even for the working process too. Like though most of the jobs nowadays have become remote, so people now just work within their comfort. But for the education processes, most people want to move into other countries to enhance their education and get a better understanding on whatever other people are doing outside there. Yeah. So yeah, we are moving into education and we are trying to with the social impact also that is unemployment, having companies come in to help their employees to move to other countries or the places that they are established in. So yeah, yeah.

Adam Becker [00:23:06]: Amalo, a couple of other questions before I let you go. One of them is, so you mentioned that the communication patterns between each of these agents is either with REST or with messages. For people that are getting started building communication protocols and interfaces between agents, are there any recommendations for resources that you would impress on them? What should they get started with?

Amolo Washington [00:23:35]: Okay, first get a better understanding on the language that you want to build on. That is the programming language that you want to do reuse. So with us we use the restful APIs because it's much more faster and at least it's not very bulk during the coding process. So it makes everything very much seamless and the performance is usually very, very fast. So that's the reason as to why. So if anyone is trying to look into building an end to end communication platform. They'll need to start from the very basic coming up to now to the advanced level.

Adam Becker [00:24:26]: Wonderful. Amalo, stick around the chat in case people have more questions. And it was absolutely wonderful having you. How do people stay in touch with you?

Amolo Washington [00:24:38]: Okay, if people have questions, I have my LinkedIn, my LinkedIn profile there. We'll just search Washington, Amolo and on X, that is at Datamart K.

Adam Becker [00:24:58]: Nice. We can put that in the chat below. Amala, where are you located?

Amolo Washington [00:25:06]: I'm in Nairobi, Kenya.

Adam Becker [00:25:08]: Nairobi, Kenya?

Amolo Washington [00:25:10]: Yeah.

Adam Becker [00:25:10]: Nice. Is that where the much of the team is located or are you guys fairly remote?

Amolo Washington [00:25:17]: The other guys in most of this founder is in US

Adam Becker [00:25:23]: Oh, yeah.

Amolo Washington [00:25:24]: Okay. Let me say half of the companies in US, then the two of the developers now in Kenya.

Adam Becker [00:25:32]: You know, there's a. I hear there's a growing scene of AI engineers and developers there. That's been pretty. That's been growing very, very fast. How is the scene there? How do you like working out of there?

Amolo Washington [00:25:48]: Yeah, just enjoy the environment and being closer to your region now, having people around you that you know mostly.

Adam Becker [00:26:00]: Wonderful. Amalo, thank you very much for joining us today. And stick around the chat and we'll see you soon.

Amolo Washington [00:26:07]: Okay, thank you.

Adam Becker [00:26:09]: Adios.

Amolo Washington [00:26:10]: Bye.

Adam Becker [00:26:10]: Bye.