I'm Mohammad Waqas, a systems engineer focused on GPU-accelerated LLM inference, observability, and local-first AI infrastructure. I build practical tooling around CUDA, llama.cpp, OpenTelemetry, distributed computing, and low-VRAM deployment.
- Building CUDA-first inference and telemetry systems for local LLMs
- Engineering distributed runtimes with MPI, TCP, async I/O, and content-addressed storage
- Connecting coding agents to local GGUF models through MCP bridges
- Operating AI workloads with Docker, Kubernetes, Helm, Prometheus, and Grafana
| Project | What it does | Stack |
|---|---|---|
| LlamaTelemetry | CUDA-first OpenTelemetry SDK for LLM inference observability and explainability | Python, OpenTelemetry, CUDA |
| CUDA NVIDIA Systems Engineering | Distributed LLM inference system with TCP networking, MPI scheduling, storage, and latency benchmarks | C++20, CUDA, MPI |
| LLM Observability Stack | Local GPU AI platform combining k3s, Ollama, Open WebUI, LangChain, and observability tooling | Kubernetes, Helm, NVIDIA |
| Windsurf llama.cpp MCP Bridge | MCP server routing coding-agent tools to a local llama.cpp server | Python, MCP, GGUF |
| CUDA MPI Llama Scheduler | Work-stealing inference scheduler with multi-rank load balancing and percentile latency analysis | CUDA, MPI, C++ |
| Ubuntu CUDA llama.cpp Executable | Prebuilt CUDA-enabled llama.cpp distribution for Ubuntu, from low-VRAM GPUs to RTX systems | Python, CUDA, llama.cpp |
Languages
AI, Inference and Observability
Infrastructure and Systems
Building efficient AI systems from GPU kernels to production telemetry.