Live speech translation and transcription in a Progressive Web App, with 29 spoken target languages, streaming audio, and copyable transcripts. Built for phones and desktops.

Direct NVFP4 MoE inference for Qwen3.8-Flash-Next on a 32 GB RTX 5090: read-only safetensors, lossless CSF decoding, RAM-resident weights, a bounded VRAM expert cache, custom CUDA W4A16 kernels, an…

Open benchmarking project for VLQ, the 24-qubit star-topology superconducting quantum computer of EuroHPC. Measures how circuit connectivity and compilation affect real execution quality using QFT,…

Technical preview: 27B ternary LLM inference on RTX 5090 with CUDA, CUTLASS/CuTe, and vLLM. Custom Tensor Core kernels, memory-efficient prefill, and MTP speculative decoding. Measured 246 .. 284 d…

AI research workspace built with FastAPI and React: hybrid BM25 + vector retrieval, inspectable citations, image understanding, and offline saved research.

Auditable quantum workload benchmarking with deterministic Qiskit workflows and CUDA CPU/GPU parity evidence

GitHub profile covering GPU systems, quantum software, and reproducible engineering

Native NIXL/UCX weight streaming for a 95 GB NVFP4 model on one 32 GB RTX 5090

TODO - add student guidance

Quantum computing IBM Qiskit Global Summer School 2026 dev week

simple pwa for weather forecast (pocasi) to read GPS and redirect to meteo pl