Skip to content

GitHub app

Optimization, reviewed in your repository

Turn a performance problem into an optimization plan, measured experiments and a pull request your team can review.

The GitHub app is available through an engagement with Geometric. Contact us to get started.

Start with the workload

A faster kernel is useful when it improves the system around it. We start with the code you want to optimize, the hardware it runs on and the behavior it needs to preserve. That gives the work a concrete target: lower latency, higher throughput or a smaller memory footprint under defined constraints.

Model and inference code

Identify the operations, model components or serving paths that matter to your workload.

Hardware-aware implementation

Evaluate changes against the target device, runtime, shapes and numerical requirements.

Measured comparisons

Compare candidates with an agreed baseline, including correctness checks and the conditions behind each result.

From issue to evidence

  1. Step 1

    Define the problem

    Describe the target in a GitHub issue: the relevant code, representative inputs, target hardware and what a useful improvement would look like.

  2. Step 2

    Approve the plan

    Review the proposed scope, baseline and evaluation criteria before optimization work begins.

  3. Step 3

    Run and compare

    Geometric's optimization systems explore candidate implementations and evaluate them against the agreed workload.

  4. Step 4

    Review the change

    Inspect the resulting pull request alongside its measured results and checks. Decide what to adopt through your team's review process.

A result you can inspect

The code change and the evidence belong together. A useful review shows what changed, which baseline was used, what was tested and where the improvement applies. Kernel benchmarks, end-to-end inference measurements and model-quality evaluations answer different questions; the evaluation should match the change.

Questions about the app

What kinds of code can we work on?

We work on model, kernel and inference code. Share your framework, runtime, target hardware and representative workload so we can assess the fit.

Will changes be merged automatically?

The workflow is built around an approved plan and a pull request for review. Your team decides whether to adopt the proposed changes.

How are improvements evaluated?

We agree the baseline and checks for the workload. Depending on the change, these can include numerical correctness, model-quality evaluations, latency, throughput and memory use.

What should I share to get started?

Start with a non-confidential description of the performance problem, your stack, target hardware and success criteria. We can then discuss the code access and data-handling requirements for the engagement. You do not need to send private source code in the first message.

Bring us a performance problem

Tell us what runs today, where it falls short and what you need to improve.