Architected the multi-agent brain behind the world's first AI-native professional network — iMessage-based orchestration with persisted memory, selective dispatch across 5 sub-agents, and a Kafka-backed event pipeline sustaining 50+ concurrent sessions.
As the Founding Software Engineer, I owned the complete lifecycle of the Franklink platform: from initial architecture to production deployment on AWS ECS. I designed an iMessage-based multi-agent orchestration layer with persisted memory, selective dispatch across 5 sub-agents (onboarding, networking, group chat networking, group chat maintenance, profile updates), and agentic RAG over 8 permissioned tools (database, email, calendar, profile, networking, group chat, conversation sync, email extraction).
The conductor agent (InteractionAgent) analyzes incoming messages and routes to specialized ExecutionAgents via a ReAct loop. Each agent accesses enriched user context pulled from Supabase profiles and a Zep knowledge graph storing conversation history and email activity analysis.
I engineered a Kafka-backed event pipeline with at-least-once delivery, idempotency keys stored in Redis (24h TTL), and a 3-tier retry strategy (30s, 2m, 10m) with dead letter queue after 6 attempts. The system sustained 50+ concurrent sessions with p99 latency within 40% of single-session baseline, using keyed concurrency and 1.5s message coalescing.
An InteractionAgent conductor selectively dispatches to 5 specialized ExecutionAgents via ReAct loops, each with dedicated tools and persisted memory through Zep knowledge graphs and Supabase.
At-least-once delivery with idempotency keys in Redis, 3-tier retry (30s/2m/10m) with DLQ after 6 attempts. Keyed concurrency and partition flow control sustain 50+ concurrent sessions with p99 latency within 40% of single-session baseline.
Agentic RAG over database, email, calendar, profile management, networking, group chat, conversation sync, and email extraction. Tools receive enriched context from Zep facts and Supabase profiles.
A stateless task-based design: the InteractionAgent (conductor) classifies intent and dispatches to specialized ExecutionAgents via Task definitions. No checkpoint serialization overhead — memory lives in Zep and Supabase. Message coalescing with a 1.5s debounce window prevents redundant agent invocations.
Three-tier memory: ExecutionMemory (per-task scratchpad), InteractionMemory (cross-task state), and TaskHistorySaver (persistent logs). Zep knowledge graph stores enriched user facts from conversation history and email activity analysis for long-term context.
Docker Compose orchestrates Kafka, workers, and the ingest service. The Kafka pipeline uses 5 topics (inbound, 3 retry tiers, DLQ) with 12 partitions and 3x replication. Deployed on AWS ECS with IAM authentication for MSK.
Walk through the full Franklink experience — from your first message to building a professional network, all through iMessage.
View Product Demo