Rust + CUDA inference engine for NVIDIA Blackwell (SM120), built to MoE models on a single GPU. Native NVFP4/FP8, hand-written kernels, explicit VRAM/RAM placement per component. On an RTX 5070 Ti it beats vLLM on decode and prefill for models that fit — and runs several vLLM OOMs on outright.

Self-calibrated EXL3 4.05 bpw quant of Qwen3.8-Flash-Next (weights on Hugging Face): calibration data generated by the model itself in coding-agent sessions, chats at four reasoning levels and on images, a measured bit split in the MoE experts, per-expert affinity-weighted Hessians. At the same size it is 12-17% closer to BF16 (KL divergence) than turboderp's 4.05 bpw quant on two independent traces. Calibration data is open under ODbL.

Open-source xray-core VPN clients (Linux · Windows · Android) with per-app/per-domain split tunneling, built on Tauri 2 + SvelteKit.

Production VPN subscription platform on Xray (VLESS + Reality) and Hysteria2, live at aegisvpn.org. Every node pulls signed config over outbound HTTPS and reconciles itself — no inbound port open on any exit server.