tweet-vibe-coded-cuda-inference

Vibe-Coded CUDA Inference Engine

Author: @ekzhang1 (Eric Zhang)
Date: 2026-04-22

Summary

“I didn’t want to wait for vLLM startup so I tried to vibe code a cuda graphed, compiled, low-latency inference engine.” Demonstrates the new reality where even CUDA kernel engineering can be AI-assisted.

Concepts

• Coding Agents

See Also

• tweet-cuda-matmul-beats-cublas — hand-tuned CUDA
• vllm-triton-attention-backend — vLLM optimizations