← All projects
2026
Automation platform backend
A fast, predictable backend for an automation platform — event-driven services, background jobs, semantic search and real-time updates in the UI.
Also available in Русский
Under NDA — names and details are generalized.
- Role
- Backend developer
- Period
- 2026
- Stack
- Python, FastAPI, NATS, PostgreSQL, Qdrant, MinIO, WebSocket, Prometheus, Grafana, Docker, GitLab CI
- p95 publish → consumer ack on the event bus
- < 10 ms
- p95 publish → consumer ack on the event bus
- tasks per minute per worker
- 1,000+
- tasks per minute per worker
The problem
A platform where users build automations that call external and AI services. The backend had to answer the UI quickly, run long jobs reliably in the background and never lose work when a downstream service fails.
Approach
- Event-driven services in async Python, talking over a message bus instead of long synchronous call chains.
- PostgreSQL for core data, MinIO for files, Qdrant for semantic search.
- Async-first API — slow work goes to background workers, results reach the UI in real time over WebSocket.
Practices
- Low publish-to-ack latency on the event bus — consumer prefetch and concurrency are sized per workload, handlers stay non-blocking and slow I/O is offloaded, so a message is picked up and acknowledged within milliseconds. Consumer lag is a first-class metric with alerts.
- Backpressure instead of overload — bounded in-flight messages, durable consumers, retries with backoff: a failing integration neither loses work nor gets hammered.
- Reusable CI/CD templates — every service inherits shared pipeline stages for linting, type checks, tests, image build and deploy, so a new service gets a complete pipeline in a few lines of config.
- Performance budgets from day one — query plans, indexes and caching are part of the design, not a fix after release.
Results
My scope is the quality and speed of the backend, not the infrastructure under load. Result: events are picked up and acknowledged in under 10 ms at p95, and a single worker processes 1,000+ tasks per minute.