infanasotku

Developing backend services and infrastructure for production RAG pipelines, improving observability, reliability, and validation of model-related changes. • Built an offline A/B testing system for full RAG trace comparison using an LLM judge, running on thousands of production-sourced requests. During an urgent generative model replacement, it identified significant regressions in 30% of client-specific pipeline configurations before rollout. • Migrated AI agents from a hardcoded execution loop to a declarative pipeline runtime based on YAML configuration and separate MCP servers. This let the ML team ship pipeline changes independently, reduced routine agent updates from the backend team’s backlog, and enabled client-specific on-premise builds through feature flags. Technologies: Python 3.12, Go, FastAPI, PostgreSQL, pgvector, S3, Redis, Celery, Kubernetes

Worked on backend infrastructure for LLM-powered products, improving request processing, service reliability, and scalability under production load. • Redesigned a synchronous LLM proxy into an asynchronous outbox-based pipeline, doubling throughput from ~300 to ~600 chat completions per minute and supporting peaks up to 1,100 RPM. Split processing into separate services coordinated through status changes and outbox messages, eliminating most 499/500 errors caused by overload and PostgreSQL connection pool exhaustion. • Introduced real-time QoS controls for 20+ LLM models, allowing the team to adjust throughput and model pressure from the admin panel without restarting services. Processor nodes picked up changes through a reconcile loop and applied concurrent request limits independently for each model. • Reworked billing and rate limiting from plain request counters to a token-bucket model inside the asynchronous LLM pipeline. This smoothed burst traffic from the largest client by 2x without reducing their limits or degrading the user experience. Technologies: Python 3.12, FastAPI, PostgreSQL, RabbitMQ, Kafka, S3, Redis, ClickHouse, Prometheus, Qdrant, Kubernetes

Led backend development for a restaurant ERP system, focusing on business modules, integrations, file infrastructure, and team delivery across a monolith-plus-microservices architecture. • Built a file management microservice that handled the full file lifecycle through presigned S3 URLs, from upload to cleanup. It also published RabbitMQ events for unconfirmed file deletion, allowing other modules, such as the knowledge base, to react to file lifecycle changes. • Led delivery of SLA, CRM, Telegram bot, file management, and knowledge base modules with a 7-person team, working across backend development, planning, demos, hiring, and production rollout. Technologies: Python 3.11, FastAPI, SQLAlchemy, PostgreSQL, RabbitMQ, Celery, Redis, S3, Kubernetes, Docker, Vue 3

Built Python backend services for a geospatial engineering application used in well drilling support, working on data parsing, geological surface conversion, computational logic, and integration with a legacy .NET desktop client. • Built an API converter for geological data, parsing the poorly documented GRDECL format and preparing it for the desktop client. • Integrated backend APIs with a legacy .NET desktop client, refactored related code, and implemented FastTree-based surface alignment logic. Technologies: Python 3.11, FastAPI, SQLAlchemy, PostgreSQL, Docker, .NET