About Maincode
Maincode is an AI research and engineering company. We build Matilda, Australia's large language model, trained, served and deployed in Melbourne. We run our own GPUs in local data centres and we build everything above them: training and serving systems, model code, evaluations, and the products that use them. We are training the next version of Matilda on it and serving the current version to the public.
About the role
You will decide, with the team, how the next Matilda is built. You will design architecture and training changes, run controlled experiments on our cluster, analyse the logs, and turn the results into decisions. You work directly with the engineers who run training.
This is an engineering role. Research here means writing the code, running the experiments, reading the logs, and writing the note that says what changed and why. If you do not want to write production code, this is not the role.
What you would actually do
Design and test architecture, objective and training-regime changes for a large language model.
Run controlled experiments at scale and identify what caused the result.
Study failure modes in reasoning, generalisation and representation.
Build evaluations that measure capability and reliability, not only benchmark scores.
Write short internal notes that turn results, including negative ones, into design decisions.
What we are looking for
Deep fundamentals. You understand transformers at the level of mechanism: attention variants, positional encodings, normalisation, optimisers, tokenisation, data mixtures, and how each affects the loss and the model's behaviour. You can diagnose a training run from its logs. You know the pre-training and evaluation literature well enough to disagree with parts of it, and you have tested some of those disagreements yourself. Production-quality Python and PyTorch or JAX is assumed.
Experience with multi-node distributed training is useful but not required.
What this role is not
Not product research. Not prompt engineering. Not fine-tuning someone else's model or building on an external API. Not a role measured by publications.
Note
Full time, Melbourne, in person. We cannot offer visa sponsorship; you need existing and unrestricted work rights in Australia.