AI & Experimental

The AI layer: from the GPU kernel to the generation pipeline.

Local LLM inference

A local inference stack in production: LLMs optimized for GPU, 262k-token context and real benchmarks. Everything that ships is measured — when a configuration doesn't hold up, I measure and report instead of accepting.

Luna — AI assistant

Local personal assistant: Qwen3.8 27B fine-tune, persistent memory, hybrid RAG (pgvector + BGE-M3 + RRF), web search via SearXNG, MCP server and a Next.js chat UI. It's the foundation of Luna V2 — the version with voice and a 3D body.

Image & video generation

ComfyUI workflows in production (SDXL and Qwen-Image): campaigns, product shots, illustration and restoration, with LoRA training for style, character or product. Generative video pipeline (MiniMax H3, WAN 2.2) with cross-shot consistency, full post-production and kernel-level debugging.

3D & creative

r6s-clone — tactical FPS

Playable Rainbow Six Siege-style demo in Babylon.js: attack/defend rounds, destructible walls, gadgets and spike. 100% procedural — zero external assets.

chess3d — 3D chess

3D chess in Three.js with full rules validation (chess.js). A single index.html, no build, no server.

Concepts

Personal projects shown as concepts only — the knowledge exists, the code is not public.

Web scraping

Extensions and automated download tools (e.g.: NFSe XML/PDF without captcha).

Reverse engineering & data analysis

Betting data analysis and probability models from real traffic.

Offensive security

Vulnerability scanner for pentesting and exploit discovery.

Voice cloning

Voice replication from short samples.