AI-Augmented Engineering Lead based in Belgrade, Serbia
24+ years building software — from safety-critical avionics certified by the FAA to cloud infrastructure handling millions of RPS at Yandex. Now pivoting full-time into AI-augmented / "vibe coding" engineering roles: shipping at LLM-accelerated pace while keeping output production-grade.
Team Lead @ AeroNAV — managing a team of 8 building an airport meteorological radar system. 11 repos, 605+ PRs; built the team's LLM-accelerated engineering process end-to-end (AI-driven PR review pipeline, parallel multi-agent refactors, RAG-enhanced review, Gitea/Redmine MCPs).
- 265 sessions, 176,848 messages, 54,759 tool calls, 18.33 B tokens, 63 active days
- 209 git commits shipped across 109 analyzed sessions, ~2,153 hours of session time
- Primary model: Claude Opus 4.6 (~73% of traffic); also Opus 4.5/4.7, Haiku 4.5, Sonnet 4.6
- 15+ projects touched: safety-critical radar, Yandex infra, LLM/ML systems, tooling, personal
Languages: Python, Go, C/C++, TypeScript, SQL
LLM tooling: Claude Code, Claude Agent SDK, Anthropic API, MCP protocol, prompt caching, multi-agent / sub-agent patterns, hooks & skills, RAG
GPU / low-level: Apple Metal (read), CUDA 12+, AMD HIP/ROCm, Strix Halo APU (dp4a + hipMalloc), MI300 tuning (aotriton flash-attn, FLA fused recurrent)
Infrastructure: Docker, Nginx, ZooKeeper, RabbitMQ, MongoDB, Redis, PostgreSQL, Elasticsearch
Cloud & CI/CD: AWS, Yandex Cloud, GitHub Actions, Gitea Actions, Jenkins
Monitoring: Grafana, Prometheus, ELK, OpenTelemetry
- 30+ avionics projects, 10+ aircraft certified by FAA (DO-178B/C)
- Large-scale distributed systems — L3/L7 load balancing, 7,000+ balancer migrations, 5 datacenters
- AI-augmented engineering at scale — production MCP server (20+ tools), LLM-driven code review pipelines on Gitea, multi-agent orchestration (5 parallel agents for cross-repo refactors, 176-commit monorepo migration in one session), full-stack project bootstrapping from zero to production
- Mentored 5+ engineers, served as Scrum Master for 3 years
- flash-moe — multi-vendor GPU port of the Metal-only inference engine that runs Qwen3.5-397B-A17B (397B-param MoE) on consumer hardware. 233 commits across four branches: CUDA (61), ROCm (46), AMD APU / Strix Halo (73, incl. dp4a + undocumented hipMalloc discoveries), and MI300 optimization (53, aotriton flash-attn + FLA fused recurrent). Article: Running 397B on a single RTX 4090.
- libfprint — new driver for FPC Disum USB fingerprint sensors (FPC1022). TLS-PSK encrypted USB protocol, SIFT-based matching for small-area sensors.
- LLVM/clang-format — merged PR: per-operator granularity for BreakBinaryOperations
- cudascope — self-hosted NVIDIA GPU monitoring (Go)
- redis-ui — modern web UI for Redis
- zookeeper-ui — web UI for Apache ZooKeeper with protobuf support
- gost-r-54084-2010 — atmospheric boundary layer models for aerospace (Python)