Andrei

Hardened Kubernetes web services and fixed memory leaks. Traced customer data across services and Airflow DAGs to deliver new features. Added retries, error propagation, and clean shutdown so services could handle network failures. Built a generic GCP Pub/Sub subscriber on Kubernetes with a Pydantic event interface, making it very easy to add new events and their processing logic. If the Pub/Sub connection died, the health probe failed and Kubernetes restarted the pod. Took a Telegram Bot prototype to production: uv and ruff, src layout, SQLAlchemy sync and async engines with connection pooling, commit/rollback, production logs formatting, Grafana dashboards and alerts. Replaced a broken in-house structlog wrapper with a custom drop-in package used across scattered repos: JSON in production, pretty-printed locally.

Maintained the customer API that served ML predictions at 50+ rps. Reworked Django ORM use across the repo to optimise database queries. When Redis was down, the cache client retried with long timeouts and held requests for 7–10 minutes. The fix was to check Redis once per request, then fall back to Postgres. Those requests then finished in a few seconds. Took a FastAPI prototype for storing and fetching ML time-series predictions to production on Kubernetes: memory leaks, database connections, autoscaling. Write P95 latency went from 3-5s to under 800ms. Added Sentry with spans on two backend services and showed the team how to read traces. Removed constructors that opened a Postgres connection and then never queried. Split customer-facing prediction data from internal-only data behind one pattern reused across endpoints, instead of copy-pasted checks. Shipped large Postgres model migrations without taking the API down. Mentored juniors on Django, ORM-efficient queries, and migrations that could still roll back. Introduced ruff for streamlined formatting and linting. Added test factories so tests started from valid production-shaped data.

Senior engineer on a team of 2-5 engineers delivering web products for Google stakeholders based in the US. The engineers were based in Europe. Took quarterly briefs, pushed back when the idea did not work, got Google security sign-off on the stack, shipped to staging and production, handled daily maintenance & monitoring. Ran the platform used by about 10,000 marketing agencies across North America, South America, Europe, Asia, and Australia. Refactored and consolidated an application spread across multiple AppEngine instances into a single AppEngine project.

Developed software products to anonymise data using K-anonymity. Built a new product based on differential privacy.