I build deep, correct systems across unfamiliar stacks — backend, AI platform, on-device NPU inference, Linux internals. AI does a lot of the typing; my job is making sure it's actually right — verified against real benchmarks, gold-sets, and byte-exact diffs, not just plausible-looking output.

• Fixed AMD's open NPU compiler so it scales to large designs — the O(n²) build walls (a kernel compile from ~9 min to ~44 s), GEMM fusion, batched operators, and a silent wrong-answer backend miscompile a hardware test caught. 11 PRs merged across mlir-aie, IRON, mlir-air and llvm-aie, 5 more in review down to upstream LLVM.
• Built a from-scratch on-device NPU inference engine for AMD Ryzen AI (XDNA2) on Linux — runs whole model graphs on the laptop NPU (a 16-model loader, 6 proven end-to-end: BERT / ViT / Whisper / Parakeet / ESM-2 …) where AMD's own answer is a closed, fixed catalog, at ~half the CPU's package power. → xdna-engine (Apache-2.0)
• Production AI for a top-3 Russian B2B accounting-software publisher: a 1С payroll-compliance agent, a self-hosted OCR→LLM document-extraction pipeline (chosen on a labeled gold-set), vLLM-on-H100.
• Smaller merged fixes upstream in zed and helm-diff; an open speculative-decoding PR to vllm.

• kde-mcp (Rust) — computer-use MCP server for the Linux desktop (KDE Plasma 6 / Wayland); drives apps through the accessibility tree, not pixel-clicks.
• tnl (Rust) — self-hosted ngrok alternative; one-command reverse tunnels behind your own Caddy + wildcard domain.
• evbridge (Rust) — evdev→Wayland input bridge for headless / nested compositors; one (small) piece of a larger, much hairier Wayland-native game-streaming system I'm building — patched Sunshine + Wolf, multi-GPU, Moonlight.
• run-script-webpack-plugin — in NestJS' official HMR docs; ~600K npm downloads/mo, 38 dependents.
• ru-bot — antispam bot for one of the largest Russian-language dev communities; in production 8+ years, ~100K users/year.
• Top contributor to rethinkdb-ts — ~63% of commits, 9+ merged PRs including releases.

• AI Platform — Model Context Protocol (3 implementations: Rust / Go / TS), agentic systems, computer-use infrastructure
• On-device & GPU inference — AMD XDNA2 NPU kernels (mlir-aie / IRON), vLLM on H100, OCR + LLM pipelines
• Backend — TypeScript / Go / Rust / Python; Kubernetes, multi-tenant SaaS, payments, high-throughput data
• Systems — Linux / Wayland / KDE deep-stack, NVIDIA driver builds, self-hosted infrastructure (Headscale, Caddy, TrueNAS)
• Leadership — tech-led teams of up to 5; hiring & interviewing, ADR-driven design, document-first handoff

📧 [email protected] · 💬 Telegram @atassis · 🌐 atassis.ru · 💼 LinkedIn