kog

Kog builds a co-designed inference stack for real-time AI agents on standard datacenter GPUs, developing model architecture, inference engines, compilers, and low-level GPU kernels. The company serves developers and organizations building AI coding agents and agentic workflows by providing infrastructure that delivers reasoning depth at product speed. On the model side, Kog has developed Laneformer 2B and Delayed Tensor Parallelism, while its systems side features the Kog Inference Engine running on standard AMD and NVIDIA datacenter GPUs to generate thousands of tokens per second per request.

The company's stated work in AI and machine learning focuses on overcoming the bottleneck of sequential generation in large language models. Kog's technology combines low-latency engines with parallel architecture, including hardware-software co-design and topology-aware algorithms that steer memory access toward optimal physical locations on the GPU. Its next major project is AGCO, an agentic compiler designed to optimize LLMs across different GPUs and optimization targets.

The open roles indicate that Kog is building a highly technical engineering team with deep expertise in compiler engineering, optimization passes, intermediate representations, lowerings, code generation, LLVM, and MLIR. The current open position is for an Agentic Compiler Engineer based in Paris, France, with remote work options available. The company's team consists of a small, high-density group of engineers, researchers, and PhD holders.