At Outlier AI, Mercor and micro1, I design evaluation tasks and rubrics that test how models reason, use tools and complete multi-step work. I also review model outputs to identify issues such as hallucination, weak reasoning and tool misuse.
On Aether, my independent AI research and engineering project, I designed five generations of agent harnesses. The current harness runs a model on software tasks in isolated Docker environments and captures traces, environment state and costs.
I built an evaluation suite covering nine capability areas, with graders checked against more than 300 known-bad solutions and simulated careless agents. I also integrated Browser Use so agents can operate web applications and verify task completion.
By analysing failed runs and public leaderboard submissions, I investigated whether errors came from the harness or the model. In 15 of 16 audited cases, the deciding information was already in the agent's environment, but the agent either missed it or overrode it with an assumption.

