Model machinery - belief state, evals, recommendations, live AI
🌍 Find Your Scene - it learns what matters to you
Tell it what draws you to a place. It asks a handful of questions - chosen just for you - and finds your cities. Watch it learn as you answer.
Adaptive preference discovery - an AI personalization research prototype, with culture as the test domain.
1 Rank what makes a place worth exploring2 Answer a few adaptive follow-ups3 Watch it learn4 Get better picks
There are no fixed questions. After your ranking, the system keeps a belief state (per-dimension estimate + uncertainty) and generates each next question for whatever it doesn't yet know that would change your ranking most - churn × uncertainty × your priorities. Tradeoffs measure values. It states its hypothesis, takes corrections, and lets behavior override answers. Rank food + history first and a nightlife question will simply never appear.
persona:
questions 0behavior events 0how sure it is about you 0%behavior weight β 0.00golden P@5 -
1 · rank what makes a place feel worth exploring (drag to order - your top 3 lead)
2 · a few questions - it chooses them for you
3 · extras (optional)
Free text matches the corpus via TF-IDF (+0.15 blend). Question selection (which dims, when to stop) is always the measurable local policy; question generation is template-composed locally, or by AI when you enable live generation (same JSON contract, weights clamped to the selector's targets, validated server-side).
what it thinks you care about (updates live as you answer)
belief state (bars = estimate, hatched = uncertainty)
embedding map
youscenegoldensaved
your places - re-rank live as it learns; ★ save what you love, ✕ skip what you don't
drag your list into order, then hit Let's go
content 0.72signals 0.08culture 0.20semantic 0.15
offline evals - 30 personas, generated-question mode: adaptive P@5 0.53 in 4 questions vs 0.33 from a 6-question fixed survey; behavior → 0.69 (click to expand)
Engine: belief state (μ per dim + σ uncertainty) · question generation over target dims chosen by churn × (0.4+0.6·uncertainty) × priority · tradeoffs for values · live profile · hypothesis with tap-to-correct (re-opens uncertainty) · behavior-vs-stated reconciliation (β = events/(events+6)) · optional AI generation with the same JSON contract. Same policy as engine.py; gates in evals_adaptive.py: P@5 0.53 in 4 adaptive questions vs 0.33 from a 6-question survey. Four layers: culture facets (what you care about) · experience style (how you explore - tradeoffs, budget, icons-vs-hidden live HERE, never top-level) · constraints (month) · behavior signals (clicks override claims).