Sriram Vivek

An event-driven payment backend using FastAPI and Apache Kafka to decouple synchronous API ingestion from asynchronous downstream processing, backed by a PostgreSQL state machine that tracks each payment through pending, processed, and failed states. Processing is idempotent — event IDs are tracked with conflict-safe inserts and Kafka offsets are committed manually, so a consumer crash can't cause duplicate charges from a redelivered event. Failed events are routed to a dedicated dead-letter queue topic with structured error context, validated by load-testing with a producer simulating 1,000 payment requests.

An end-to-end RAG pipeline over Anthropic's API documentation — markdown-aware chunking into roughly 3,500 chunks and a self-hosted Qdrant vector store, with full ingestion (including a first-time ~550MB embedding-model download) taking about 24 minutes on a Hetzner AX41 dedicated server. Retrieval is exposed through a custom MCP server (built on the Python MCP SDK) with 4 tools, wired into an agent framework so the agent can only retrieve through MCP — no direct database access. The whole stack runs on a self-managed VPS behind Nginx/TLS with Docker, systemd, and a scripted GitHub Actions redeploy, with a Next.js frontend that includes a retrieval-inspector UI showing citation sources and similarity scores.

A multi-service gRPC backend with a central orchestrator aggregating responses from separate User and Search services over protobuf-defined contracts, using server-side streaming for real-time result delivery. Circuit breakers and exponential backoff retries contain partial failures across services, and mutual TLS secures every cross-service call. Distributed tracing runs through OpenTelemetry gRPC interceptors capturing both function- and business-level spans, visualized in Jaeger, with Prometheus and Grafana for metrics — the observability stack runs via Docker Compose.