Pavel Savinov
Senior Backend Engineer (Go)
I build production Go services: search-results APIs, asynchronous logistics workflows and promotion systems. My work at Aviasales, Uzum and Avito spans performance profiling, PostgreSQL data modelling, Kafka, transactional outbox, SLOs and incident response.
Based in Belgrade, Serbia. Open to senior Go/backend roles, remote work, B2B contracts and relocation with work-permit sponsorship.
Production experience
Uzum Market (Uzum Technologies)
Backend for the marketplace's parcel-sorting warehouses. Go, PostgreSQL and asynchronous workflows over Kafka.
- Took phase 2 of the sorting-center program from PRD to production in 4 months: domain and PostgreSQL data model, Kafka integration via transactional outbox, and most of the service implementation. Live at the first sorting center; rollout to approximately 10 centers was planned.
- Designed per-item routing over a graph of sorting walls and cells, with topology checks and load balancing across duplicate walls.
- Encoded sorting rules as domain invariants in a dedicated stage-transition module and defined SLOs for core operations.
Aviasales
Search-results API for a flight-search platform.
- Improved search-results API latency through pprof-driven profiling, caching and request shaping.
- Freed approximately 1 GB of RAM per instance (approximately 30%) across hundreds of instances by fixing memory leaks and retuning caches.
- Designed and shipped config-driven badges with the expr expression language. Moved badge logic from iOS, Android and web clients to the backend; rollout went from a sprint plus mobile releases to a 30-minute config change.
- Wrote a search-results storage design from access-pattern analysis and load projections across three data centers. It started the redesign; the Search team implemented its own version.
Avito
Paid promotions and Billing integration. Go and PHP.
- Owned SLOs, alerting, tail-latency and timeout fixes across 20+ promotion microservices, keeping 99.9%+ availability at approximately 3.5k RPS.
- Migrated a Redis cache serving approximately 3.5k RPS from a single instance to a sharded cluster with a phased rollout.
- Designed a feature-flagged cutover from legacy promotions to a new architecture, with fail-safes and rollback; delivered Billing integration for a new paid feature.
Far Eastern Federal University
Taught Databases (PostgreSQL) and Web Programming to 50+ students.
Skills
- Backend engineering
- Go (Golang), concurrency, pprof, SQL, Python, REST, gRPC, Protobuf
- Data and distributed systems
- PostgreSQL, Kafka, Redis Cluster, transactional outbox, idempotent consumers, event-driven architecture
- Reliability and delivery
- SLOs and SLIs, Prometheus, Grafana, Jaeger, OpenTelemetry, incident response, Docker, Kubernetes, CI/CD
- AI-agent tooling (personal projects)
- Claude Code, Codex, MCP, agent skills, LLM APIs, structured output
Public projects
Personal projects and open-source work, separate from my production experience.
- pr-review - An agent skill that reviews a pull request by running it: three builds, one subtraction, then probes for behavior that tests do not cover.
- ProtonDB compatibility prediction - A two-stage LightGBM classifier predicting Linux game compatibility from 350K+ community reports. Built with Claude Code as a research partner; per-game F1 of 0.871.
- Graph LLM Chat - A local-first branching chat UI with streaming, token accounting, summarization at context limits and multi-server MCP integration.
- PHP: The Right Way, Russian translation - Russian translation of PHP: The Right Way, maintained by the community.
Engineering writing
- A Recruiter Sent Me a Take-Home Test. It Takes Its Orders From an Ethereum Wallet.
A fake recruiter, a VS Code extension that shows one script and runs another, and malware that looks up its server in Ethereum transactions. Read without running a line of it.
- I Built the Detector Before the Protocol. Then Two Agents Fought Over It.
Adversarial development with Claude: a lab, a battery of ML classifiers as the referee, a zoo of every past detector, and five rounds of measurement mistakes that taught me more than the wins.
- I Don't Know ML. Claude Does. 0.871 F1 on Predicting Linux Game Compatibility.
Building a system that predicts Linux game compatibility using 350K+ community reports and Claude as an ML research partner.
Education and languages
- Master's degree, Applied Mathematics and Informatics, Far Eastern Federal University, 2017-2019.
- Bachelor's degree, Applied Mathematics and Informatics, Far Eastern Federal University, 2013-2017.
English: C1. Russian: Native.
Let's talk
Working on Go backends, search infrastructure or reliable event-driven systems? Email me with the role or the problem you're working on.