Models trained to write better kernels.
Our hardware-specific LLMs power AI agents that find bottlenecks in live inference and write hardware kernels to address them.
Production traffic guides every improvement.
The loop starts with live inference. Our agents use what they observe to write better kernels, test their changes, and put verified improvements back into serving.
View Mosaic benchmarksObserve real workloads
Execution graphs, context lengths, and changing batch sizes show where inference spends time.
Write better kernels
Our AI agents use hardware-specific LLMs to generate and refine code for those bottlenecks.
Test against the workload
Check correctness and measure performance under the conditions that exposed the problem.
Deploy and measure
Release verified improvements into Geodd’s inference service, then monitor their effect on production performance.
Feed the next cycle
Successful kernels and measured results inform the next experiments and further fine-tuning of our hardware-specific LLMs.
The models powering
this process.
Mosaic, Druze, and Strata power the agents in this process. Each is fine-tuned for the hardware it optimizes.
Druze · AMD
Kernel development for AMD GPUs and ROCm.
Coming soon
Strata · Tenstorrent
Kernel development for Tenstorrent accelerators and TT-Metalium.
Coming soon
Run optimized kernels on your own GPU.
Our autonomous AI agents use Mosaic to develop model-specific runtime kernels. Free Docker images package these kernels with vLLM, SGLang, or TensorRT-LLM so you can test their performance and use them in your own projects.
Use the improvements through one API.
Verified kernel improvements feed back into Geodd’s inference service. Run the models you need through our API while we continue improving how they execute.