I run the IT infrastructure @ EIFER - European Institute for Energy Research and build AI tools on the side.

A from-scratch C++/CUDA inference engine that targets exactly one architecture: the NVIDIA RTX 5090 / RTX PRO 6000 (sm_120a). No portability layer, no wrapper around llama.cpp or vLLM. It ships its own GGUF & SafeTensors loaders, BPE tokenizer, paged KV cache, attention/MoE kernels, Gated DeltaNet + Mamba2 scans, CUDA Graphs, and an OpenAI/Anthropic-compatible server built for agentic AI: tool calling, json_schema constrained decoding, separable reasoning channels and concurrent sub-agent fan-out.

The fastest single-GPU backend on the 5090, and every line of its ~97k is written by Claude Code.

• ⚡ GGUF dense decode: Qwen3-8B Q8_0 at ~270 tok/s, +37-72% over llama.cpp
• 🚀 NVFP4 30B-class MoE decode: ~257-338 tok/s, uncontested on sm_120 (vLLM's NVFP4 path needs tcgen05, absent on consumer Blackwell)
• 🎯 NVFP4 long-context prefill: at-or-ahead of vLLM (MoE pp4096 +4%, pp2048 +27%, TTFT wins everywhere)

Every number measured, dated and commit-anchored: github.com/kekzl/imp

• PromptMill: AI prompt generator for video & image models (Wan2.1, FLUX, SD, Midjourney). Local LLMs with GPU auto-detection.
• entra-id-secrets-notification: Monitor Entra ID secrets & certificates for expiration. Alerts via Teams, Slack, Email.
• mailcow-ai-filter: AI-powered email sorting for MailCow. Auto-generates Sieve filters using Claude or local LLMs.

Infrastructure · VMware vSphere NetApp OpenZFS Linux Docker Ansible Entra ID Microsoft 365

AI / ML · CUDA C++20 LLM Inference LLM Pretraining Claude Code

Custom Water Cooling · PC Building · Home Automation · AI Video Generation · Stock Market · Skiing