Computer Engineer & AI Developer. I work on AI inference, model serving, and agent systems — building tools that bridge research and practical deployment. I also publish technical reports on LLM inference and AI security.
- AI Inference & Quantization — GGUF quantization of MoE language models for local inference on consumer hardware, and research on inference acceleration (speculative decoding, multi-token prediction)
- Agent Systems — autonomous AI agents for server management, deployment automation, and multi-agent workflows; Claude Code on alternative model backends
- AI Security — trust boundaries and supply-chain integrity of open-weight and air-gapped LLM deployments
- Systems Engineering — self-hosted infrastructure, network security, privacy tooling on Linux and macOS
- Speculation or Substance? Rethinking How LLMs Get Faster (2026) — speculative decoding vs. multi-token prediction: architecture, cost, energy, and the real bottlenecks of production inference · DOI
- Can We Trust Open-Weight Large Language Models? (2026) — a security perspective on air-gapped deployments, supply-chain trust, and the limits of transformer-based systems · DOI
AI & LLMs
- claude-code-deepseek — Claude Code on DeepSeek V4 through its Anthropic-compatible API with the full 1M context, plus a sourced cost comparison against Claude plans
- anthropic-openai-proxy — local Anthropic ↔ OpenAI API translation proxy that lets Claude Code drive vLLM, SGLang, or any OpenAI-compatible backend (thinking, tool calls, SSE streaming)
- Stack-in-a-Click — one-click SaaS provisioning with Hermes Agent + Stripe Projects, built for the Hermes Agent Accelerated Business Hackathon (NVIDIA × Stripe × Nous Research)
- SRMS — AI-powered infrastructure issue reporting: a single photo goes through a two-stage AI pipeline and is routed to the responsible unit. Graduation project with Team Vexra at Atılım University; its DistilBERT classifier (ONNX Runtime) reaches 90.7% accuracy over 14 categories on a 300-sample benchmark
Open Source
- Deep-Live-Cam — AMD DirectML optimizations (GPU face detection, detection throttling, pre-load fix), merged upstream (PR #1710)
Systems & Tools
- Ultimate-Privacy-Server — Pi-hole + Unbound + PiVPN (WireGuard) + Zapret DPI bypass on a single Linux server, with a step-by-step guide in English, Chinese, and Turkish
- PowerFlow — live power-flow Sankey diagram for Apple Silicon Macs (adapter → battery → system → CPU/GPU), no sudo required
- Simple-Metadata-Remover — Windows app that strips date, location, and device metadata from videos with FFmpeg, without re-encoding