We build capable AI models with small compute. ⚡
Frontier AI increasingly assumes frontier-scale resources. We work in the opposite direction: getting strong reasoning, forecasting, and agentic behaviour out of models small enough to train, audit, and deploy without a data centre. We care about what a model remembers, how it reasons, and how little compute it can do both with.
The group is based at the Applied Artificial Intelligence Institute (A²I²), Deakin University, Australia.
- 🧠 Memory & long-term dependency. How models retain, retrieve, and revise information beyond their immediate input — from internal state to explicit, retrievable, and agent-maintained stores.
- 💸 Efficient reasoning under a compute budget. Getting more capability per FLOP: allocating test-time compute where it matters, and small models that reason above their weight.
- 🗣️ Multi-agent systems. How agents debate, exchange information, and stay diverse rather than collapsing to a single voice.
- 📈 Time-series intelligence. Forecasting, error correction, and memory for temporal and streaming data.
