We build infrastructure for AI scientists: systems that help them learn from prior work, run better experiments, and produce research that others can audit, reproduce, and build on.
Our open-source work spans the full loop: preserve what an AI scientist learned, put new knowledge into the model, steer the next experiment, and make the whole process inspectable.
| Project | Why it matters |
|---|---|
| Agent-Native Research Artifact | The memory and trust layer for AI science. It turns claims, evidence, code, decisions, and dead ends into research that other agents can verify and build on. |
| PULP-LLM | Turns an entire chip platform—register maps, RTL, drivers, build system, and issue history—into model knowledge. The resulting 9B open model beats Claude Opus 5 on closed-book chip questions (81.6% vs. 72.2%). |
| fuzz-testing-for-science | Gives autonomous scientists the feedback they are missing: where a theory breaks, what remains unexplored, and whether a claimed discovery survives its own evidence. |
| ara-cli | Turns ARA from a file format into working infrastructure: one local binary validates the research, renders its full trajectory, and enforces the format in CI. |
Explore our work at agenticresearch.sh or browse all public repositories.