DeepLumen is a multi-tenant AI brand analytics SaaS platform I architected and delivered end-to-end at Jancsitech, covering a React frontend, backend APIs, and database design.
Architecture
The backend follows a layered architecture (API → Service → Repository) with Temporal workflow orchestration for long-running tasks and a gRPC microservice with RabbitMQ async processing. PostgreSQL schemas use tenant-scoped isolation with indexed access patterns, supporting efficient query performance and safe schema evolution.
Reliability
Migrated async processing from RabbitMQ to Temporal, simplifying long-running task orchestration and replacing custom retry/DLQ logic with built-in reliability guarantees. Implemented idempotent task handlers, retry strategies, and dead-letter queues (DLQ) to ensure reliability under failure scenarios.
AI & LLM Pipelines
Built production LLM pipelines integrating multiple providers (OpenAI, Gemini), including RAG workflows and multimodal analysis (audio/video/image) of websites to deliver actionable brand optimization insights. Async orchestration handles rate limits and failures gracefully.
Observability
Implemented observability stack with structured logging, metrics (latency, error rate, queue depth), and alerting for early detection of system issues.