AI Researcher

About the Role

Maincode develops foundational AI models entirely on Australian infrastructure, managing every step from architectural design to deployment. We have created Matilda, the first large language model built and trained from scratch in Australia, and are currently scaling to its next iteration with a new compute cluster now operational for public use.

We seek AI researchers to contribute to the core architecture, training processes, and evaluation strategies of large-scale language models powering Matilda. This role emphasizes hands-on work beyond incremental benchmarks or paper publications, collaborating directly with engineering teams to design models that learn efficiently and behave reliably in production environments.

Key Responsibilities

• Devise and assess architectural innovations and training procedures for large language models.
• Conduct large-scale controlled experiments to isolate causal factors.
• Investigate failure modes relating to reasoning, generalization, robustness, and representation.
• Shape training objectives, data mixes, and optimization techniques to influence model behavior.
• Create and refine evaluation methods focusing on model capabilities and reliability beyond simple scoring.
• Analyze training dynamics through logs, metrics, and model outputs.
• Partner with machine learning systems engineers on distributed training operations.
• Document experimental outcomes clearly for informed design decision-making.

Candidate Profile

We value deep reasoning skills, experimental rigor, and the capacity to progress amid ambiguity.

• Proficient in writing and executing production-level machine learning or research code.
• Strong Python programming skills along with experience in PyTorch or JAX frameworks.
• Well-versed in transformer-based language models, including fundamentals of pre-training and evaluation.
• Skilled at experiment design, result interpretation, and clear communication of tradeoffs.
• Ability to operate closely with infrastructure constraints and performance considerations.
• Interest or exposure to reasoning-centric architectures and training methods that extend beyond conventional approaches and standard large language models.

Additional Desired Skills

• Experience with distributed training concepts and tools such as data parallelism, tensor parallelism, sharding, and checkpointing.
• Familiarity managing extended training across multiple nodes.
• Knowledge of large model training systems and frameworks (e.g., Megatron-style, DeepSpeed-like tools, FDSP).
• Comfort across full workflows including training, evaluation, and deployment challenges.
• Experience in ROCm-based computing environments.

Work Approach

• Engage in hands-on research using coding as a core tool for innovation.
• Swiftly and accurately transition between theoretical concepts and their implementation.
• Favor rigorously controlled experiments over wide exploratory runs.
• Leverage logs, metrics, and observed model behaviors to inform design choices.
• Collaborate closely with engineering teams to scale and test ideas effectively.

Role Clarifications

• This is not a product research or prompt engineering position.
• It does not involve fine-tuning existing models or wrapping APIs.
• The focus is exclusively on building and improving Matilda, a model trained and developed internally.

Why Maincode?

Maincode operates the entire technology stack including training infrastructure, model architecture, evaluation systems, and deployment platforms. We manage one of Australia's largest private AI compute environments dedicated to the end-to-end lifecycle of large-scale model development and deployment.

This is a unique opportunity to directly engage with building a large language model from scratch inside Australia.

Additional Information

• This full-time position is based in Melbourne, requiring on-site presence and close teamwork with our in-person staff.
• Currently, visa sponsorship is not offered; candidates must have existing unrestricted work rights in Australia.