Architected the multi-agent brain behind the first AI-native professional network, from zero to one.
As the Founding Software Engineer, I took charge of the complete lifecycle of Franklink platform: from the initial architectural diagrams to the production deployment on AWS. My core responsibility was translating the product vision into a tangible technical reality, selecting the right stack to balance rapid iteration with long-term stability.
I co-designed a Planner-Worker multi-agent architecture that moves beyond simple chatbots. By decoupling "reasoning" (understanding intent) from "execution" (scheduling, database lookups), we created a robust system that could handle complex, multi-step networking requests with high reliability and verifiable state transitions.
Beyond just the AI logic, I focused heavily on reliability and scale. I personally engineered the Kafka-based concurrency control mechanism that ensures thread-safety across distributed agents. I also pioneered a safe tool execution layer that ensures the AI handles sensitive user data (emails, calendars) securely.
Co-architected a system where a central "Reasoning Agent" delegates specialized sub-tasks to focused "Worker Agents", enabling complex problem-solving that a single model couldn't handle.
Engineered a distributed backbone designed for massive scale. I implemented a custom Kafka protocol with strict idempotency keys and retry policies, ensuring reliable, exactly-once processing even under high concurrent load.
Designed the complete Docker and AWS ECS testing and deployment workflow. This containerized pipeline streamlined testing and production releases, allowing the team to iterate rapidly with confidence.
The system is built on a graph-based orchestration framework. This allows for validatible state transitions and "human-in-the-loop" checkpoints, ensuring that the AI operates within safe bounds while autonomy is maximized.
We designed a state management layer that persists user context across sessions. This enables the agents to cooperate with short-term and long-term memory, understanding user's evolving demands and value propositions over months of engagement.
The platform runs on containerized microservices. My key contribution was the Kafka event pipeline which decoupled agent thought processes from user I/O, allowing for independent scaling of reasoning nodes.
Walk through the full Franklink experience — from your first message to building a professional network, all through iMessage.
View Product Demo