Sub-millisecond task queue
A C++ / CUDA extension for latency-sensitive inference primitives, benchmarked against the PyTorch reference path.
0.008 ms median GEMV, 4.5x average speedup on a GTX 1650.
cat cuda-queue.md
ls links/
- panea/Rust GPU terminal runtime. WGPU, ConPTY/PTY, fuzzed VT parser.
- wmux/Multiplexer where the server owns the session. Clients come and go.
- geodesics/Light around a black hole. C++ RK4 compiled to WebAssembly.
- cuda-queue/Inference primitives in CUDA C++, 4.5x average speedup over the PyTorch path.
- cuda-task-queue.md2025-07-20 · Building High-Performance CUDA Kernels for Low-Latency ML Inference
- sentiment-pipeline.md2025-07-20 · Building a Production-Ready Sentiment Analysis Pipeline
- cnn-vs-resnet18.md2025-07-18 · Custom CNN vs ResNet18 (No Pretraining) on CIFAR-10
- fraud-detection.md2025-07-13 · Building a Production-Ready Fraud Detection System
- Feb 2026 – now
- Founding Engineer · Prescient Labs
- Aug 2024 – Jun 2025
- Founding Engineer · NeuralThread AI (acquired)
Shreshth Kapai
I like AI, systems and math.
shreshth@london
--------------
os the web
shell this one
theme catppuccin mocha
font jetbrains mono
uptime
type cd .. ⏎ to go back · /help for moretap cd .. at the bottom to go back