Start a project

Service

Faster models, smaller bills

Compiler passes, custom kernels and profiling from the model graph down to the silicon — so inference and training run faster on the hardware you already pay for.

A good fit if…

  • Inference latency or GPU cost is limiting what you can ship.
  • You are bringing models to new or custom accelerators.
  • Hot paths in your native code need serious performance work.

What we do

Profiling & roofline analysis

Trace from framework to kernel to find whether you are compute-, memory- or launch-bound.

Graph compilation

Fusion, layout and quantisation passes with torch.compile, XLA or MLIR-based stacks.

Custom kernels

Triton or CUDA kernels for the operations that dominate your profile.

Serving optimisation

Batching, KV-cache management and speculative decoding for LLM inference.

Accelerator bring-up

MLIR / LLVM lowering for new hardware targets.

Regression guards

Benchmarks and performance budgets in CI so gains don’t quietly erode.

What you get

  • Profile report with ranked opportunities
  • Optimised kernels or compiler passes
  • Before/after benchmarks
  • CI performance budgets

How an engagement runs

  1. Diagnose (1–2 weeks). We read the code and the data, and agree the baseline and the number we expect to move.
  2. Prove (3–5 weeks). A thin, real slice in production, measured against that baseline.
  3. Scale and hand over. Widen the change with your engineers, then leave docs, runbooks and a team that owns it.