Gamgee Labs

Inference is too slow and expensive. We rewrite the kernels that make it faster and cheaper.

We're a small team writing custom GPU kernels for the models teams actually run in production.

Inference speed and cost are decided in the kernel.

The kernels that run your model on the GPU — below the serving framework — set how fast each token comes back and what it costs. Off-the-shelf kernels are built for the general case, not your model. We rewrite them for the model you actually run, so the same workload comes back faster and cheaper — on your own GPUs or a managed cloud.

Everyone else tunes libraries. We write the kernel.

The usual approach

Tuning libraries

Take vLLM, TensorRT, or an off-the-shelf serving library and turn the knobs across flags, batching, config against the workloads they anticipate or observe from their clients. It's bounded by what the library already knows how to do.

Our approach

Custom kernels

We write the kernel for your specific model and your specific hardware. That's the difference between tuning what's already there and owning the layer where the cost and latency actually lives.

Stage

Optimized kernels for Gemma 4B.

We're building a custom kernel path for the 5.1-billion-parameter Gemma 4 model — a serving stack that runs it faster and cheaper than the stock path, kernel by kernel.

Who we are

Writing kernels at this level takes a rare kind of systems discipline — counting nanoseconds, defending the worst case, owning the machine down to the cache line. It's the discipline of low-latency trading, where a microsecond of jitter loses the trade, and it's where our team spent years before turning it on inference.

Working on inference cost you can't get past?

We're taking on a small number of early partners. If your own models are getting expensive to serve, let's talk about what a custom kernel path would change.

Book a call