Research
Research
How we make agents better at research, and how we put strategies on real hardware.
Agents that do quant research
Our research system runs many agents in parallel. They decode and validate raw market data, build backtests and search for strategies. Infrastructure built for one market is reused for the next, so each new market costs far less than the first.
We also study the research process itself. Agents that improve their own research methods produce better risk-adjusted out-of-sample results on the same candidate-evaluation budget.
- EvoQuantLab. Automated quant research with massively parallel AI agents. arXiv, 2026.
Hardware
Low-latency trading runs on FPGAs. Agents automate the design, deployment and testing of candidate designs on real boards, and measurements drive the next iteration. At the core is our automated model-to-RTL toolchain, which grew out of our academic work on compiling models and optimisation solvers to hardware.
- RSQP: problem-specific architectural customization for accelerated convex quadratic optimization. ISCA 2023.
- Multi-issue butterfly architecture for sparse convex quadratic programming. MICRO 2024.
- FPGA optimisation-solver research in collaboration with Stephen Boyd (Stanford).
A full publication list is on Maolin Wang’s site.