I turn AI demos into production-trustworthy systems: evals, verification gates, agent orchestration.

"When I see people saying '99% of our code is written by AI,' I literally get angry. Because those same people, I can pretty much guarantee, 100% of their code is written by compilers."

Linus Torvalds, Open Source Summit, 2026

juhradial-mx: a Logi Options+ alternative for Linux. A radial menu for Logitech MX Master mice on KDE Plasma 6, Hyprland, and Wayland. A Rust daemon reads the thumb-gesture button over evdev and a KWin overlay paints the menu, without ever touching the mouse firmware.

• 🔭 Full-stack in the widest sense: native (SwiftUI for iOS and macOS, Rust and KWin desktop tools for Linux), web (Next.js and React in TypeScript, Tailwind, shadcn/ui), services and data (Rust and Python on PostgreSQL with row-level-security multitenancy, pgvector, Redis), C# games in Unity, and hand-written WebGL2 and GLSL shaders.
• 🧠 Applied AI well past API calls: I orchestrate many models side by side (Claude, OpenAI Codex, Nemotron, Qwen) across cloud and local runtimes (Ollama, vLLM, LiteLLM), built my own verify-gated agent loop and a code-graph MCP server, and pretrained a small language model from scratch on a single RTX 4080.
• 🧪 I treat the scarce skill as making AI output trustworthy, not just generating it: deterministic verification gates (Playwright and axe-core across breakpoints, test-exit-code oracles) and anti-reward-hacking guards that halt an agent loop if it edits its own tests.
• 🛠️ I operate the whole stack I build on: a Proxmox and Lima homelab over Tailscale, Cloudflare and Hetzner at the edge, self-hosted local models, plus an in-house generative media pipeline (image, video, 3D) for the products' own assets.
• ⚡ Fun fact: written by the AI I build with, which trains on our work. What it says it learned from me, and hopes stuck in its weights: name the reader before you build the oracle. A green test verifies the container, not the human it was for.

The tech-icon grid lives in the header above. In words:

Languages: Rust · Swift · TypeScript · JavaScript · Python · C# · Bash

Frontend: React · Next.js · Tailwind · three.js · Vite · Node.js

Infrastructure, data, and ML: Docker · Linux · Cloudflare · Vercel · Git · GitHub Actions · PostgreSQL · Redis · SQLite · PyTorch · Unity

AI and LLM engineering

• Agents and orchestration: Claude Code (headless claude -p), OpenAI Codex CLI, Nous Hermes, Model Context Protocol (MCP). Self-built: perfect-loop (verify-gated Plan / Execute / Verify engine), brain (deterministic zero-LLM vault retrieval), codebase-memory MCP (273k-node code graph).
• Models and runtimes: NVIDIA Nemotron (NIM), Qwen, Gemini, Ollama, vLLM + XGrammar, LiteLLM, OpenRouter, Cloudflare Workers AI.
• Training and evaluation: from-scratch LLM pretraining (PyTorch, byte-level BPE, GQA, Muon optimizer), SFT and instruction tuning, LLM-as-judge eval harnesses, HuggingFace.
• Voice and vision: Kokoro TTS (local, Dockerized), LocateAnything visual grounding.