Independent research · Singapore
Red Cat Labs develops methods for automated experimentation and algorithm discovery, and applies them to demanding computational workloads. Current work focuses on evolutionary search, GPU kernel optimization, and systems that learn to conduct better experiments.
Led by Martin Andrews in Singapore, we work with hardware and software teams through focused research engagements and longer-term sponsorships.
Published work
An LLM-driven framework for iterative GPU kernel optimization. The system selects previous implementations, proposes experiments, and evaluates new kernels through an evolutionary process. The published work explores optimization for AMD MI300 hardware with limited feedback.
Learning to search across multiple turns. This work investigates reinforcement learning for document-search agents and the value of longer interaction horizons, evaluated on a legal document search benchmark.
Research focus
An automated researcher must decide what to try next, which previous discoveries to build on, and when a promising result needs more verification. We investigate those decisions through computational problems with measurable outcomes.
The aim is to find useful improvements within practical limits on compute, time, and experimental access.
Investigating how to make an important workload run effectively on a target accelerator, including kernel generation and optimization under unfamiliar hardware constraints.
Testing whether automated experimental search can improve a costly numerical routine or performance-critical workload while preserving correctness.
Studying selection, memory, experimental design, and verification to make computational search more effective within a fixed budget.
Work with us
Establish a baseline, investigate the opportunity, and determine whether a larger optimization effort is worthwhile.
Apply experimental search to an agreed workload and hardware target, delivering validated implementations, measurements, and integration guidance.
Support a sustained programme around related problems, with agreed milestones and opportunities to develop reusable methods and publish research.
Each engagement begins with a defined problem, evaluation criteria, and an experimental budget. Work is scoped around a useful research outcome, including a clear decision when further optimization is unlikely to pay.
Start a conversation
Send a short description of the workload, the current bottleneck, and what an improvement would enable. If known, include the hardware or software environment and the timescale. We can then discuss whether a focused research engagement would be useful.
Looking for courses? Visit Red Dragon AI training.