Rebuilding Search: Velocity Unlocks Correctness, Not the Other Way Around (www.sebastiansigl.com)

A year-long teardown and rebuild of a production search core reframes search as a data- and product-first problem rather than merely an algorithm or infrastructure challenge. The team discarded five common assumptions: that search is primarily an engineering problem, that correctness must be engineered up-front, that offline metrics are the ultimate signal, that rigid functional roles are optimal, and that technical elegance alone guarantees product success. Instead they found that high-quality, near-real-time user signals, tight feedback loops, and fast iteration drive real relevance far more than complex models trained on stale offline data. Technically, the practical lessons are clear: prioritize pipelines that minimize time from user action to logging, retraining, and deployment; favor simple ranking models trained on fresh online signals; use scrappy end-to-end production experiments and A/B tests as the primary validation; and tie every change to business KPIs (engagement, retention, revenue) rather than nDCG or other offline metrics alone. Organizationally, enable cross-functional tooling so data scientists can run experiments and engineers can observe user behavior, then harden systems only after validated impact. The implication for ML/MLops teams is a shift toward velocity-driven correctness, pragmatic observability, cost-aware tradeoffs, and a product mindset that aligns technical work with measurable user and business outcomes.