PRD == Evals: How AI Blurs the Product–ML Engineer Boundary

Albina Munirova combines ML engineering and product leadership: requirements for a probabilistic product cannot be separated from evidence of behavior. Model APIs and coding agents let product managers prototype and engineers integrate existing models; the roles converge around acceptable outcomes.

PRD == evals does not replace a document with questions. An eval stores scenarios, criteria, prohibitions, regressions, and a release threshold. For an investment RAG assistant, the team samples production pairs, groups failures, and keeps fixes in regression. A perfect static score likely signals overfitting.

The loop needs domain judgment: tests check structure, LLM-as-a-judge checks meaning, and specialists handle difficult rules. Run frequency follows the cost of people, inference, and failure; guardrails reduce risk but add latency. An MVP without data, routing, security, and a quality owner is not production.

A general assistant raises expectations: Oleg led T-Bank to domain systems with separate evals for investors, children, and travelers. An AI Product Builder owns the path from need to improvement without replacing integration experts. In a five- or six-person team, product learns evals while ML learns users and economics.