The Brave browser doesn't have a persistent downloads button (iykyk), so I vibecoded one. It adds a toolbar button that never disappears, so you can open, reveal, or drag-and-drop your recent downloads.
We re-implemented GPT-2 inference in CUDA and optimized it for NVIDIA A40 GPUs, using Nsight Systems and Nsight Compute to find bottlenecks. We added FlashAttention-2, sliding-window attention, KV caching, and CUTLASS kernel fusion. Inside the matmul and attention kernels, we hid memory latency with double-buffering and avoided bank conflicts by swizzling shared memory.
We built a real-time video pipeline on a Spartan-7 FPGA that captures a live OV7670 camera feed and renders it to an HDMI monitor. Our SCCB controller configures the camera to output a quantized 12-bit RGB444 QVGA stream, small enough to hold a full frame in on-chip BRAM, and a dual-port frame buffer bridges the camera and display clock domains. On the display side, we generate 640×480 video timing, upscale each frame, and composite color filters, text, and a sprite pixel by pixel, all in 813 LUTs at 0.35 W.