Most rec systems ask you to rate 20 things before they say anything useful. We didn't think that was acceptable, so X-Brain ships with a hybrid cold-start: a 15-question adaptive onboarding plus cohort assignment plus a content-based scorer.
Onboarding (15 pairwise comparisons). Each question is "would you rather have product A or product B?" and the choices are picked to maximize information gain across feature space. After 15 picks, we fit a tiny ridge regression — your starter preference vector.
Cohort assignment. K-means over the population's preference vectors. We assign you to your nearest cluster (k≈12 per tenant). Your "day-one" predictions use the cohort's pretrained model, which is way better than nothing.
Content-based scorer. A cosine-similarity floor: items similar to what you liked in onboarding score higher. Blended into the final prediction with a weight that decays as your personal data grows.
Final blend. α·personal + β·cohort + γ·content where (α, β, γ) is a function of your rating count. New users start at α≈0; users with 50+ ratings have α≈1 and the priors fade.
The result: day-one predictions feel personal, not generic. By day 30 they're uncannily accurate.