Large language models are increasingly used in social-science and market research to stand in for human respondents, exploring questions ahead of fieldwork or piloting instruments. Existing systems address one facet at a time (a population’s demographics, an individual’s profile, or an answer distribution elicited from a single prompt), and few serve users without social-science training. We present Alyas, a research-simulation platform in which an agentic methodologist assistant guides the user from a research goal to a study design, laid out as a branching graph of research artifacts, and simulates its field phase over a synthetic population grounded in institutional statistical sources. On held-out Italian and US microdata, the generated populations reconstruct demographic structure the engine was not given, and, on the Italian microdata, profile-conditioned respondents answer close to the real subgroups they simulate.