Curate
The system learned from post-match feedback instead of treating the first stated preference as permanent truth.
Curate combined structured profile data, conversational notes, visual preferences, and match outcomes.
The platform supported matchmaking judgment rather than replacing the human relationship work.
The approach applies to any product where preferences evolve and simple filters miss the real decision.
Outcomes are based on anonymized implementation records. Client identity, user profiles, match data, and sensitive preference details are withheld.
Overview
How Curate turned subjective preference into a learning recommendation system
Curate was a matchmaking platform with a problem every recommendation product eventually faces: people do not always know how to explain what they want until they react to what they do not want.
The work was not just about dating. It was about building a preference intelligence layer: a system that learns from feedback, context, and outcomes without reducing human taste to a static filter.
| Project context | Public detail |
|---|---|
| Client profile | Matchmaking technology platform, anonymized under NDA |
| Implementation scope | Preference modeling, feedback loops, profile context, and recommendation support |
| Engagement length | Twelve-week implementation from model design through product integration |
| Primary challenge | Users could describe preferences imperfectly, while match quality depended on signals that changed after real interactions |
| Broader relevance | Any business matching people, clients, products, candidates, homes, or services can face the same preference gap |
| Private data | User profiles, match results, conversation notes, images, and sensitive preference details withheld |
The Problem
Why people rarely know how to describe what they want
Most recommendation systems start by asking users what they prefer. That is useful, but incomplete. Human preference is often revealed through reaction, comparison, hesitation, and repeated feedback.
Curate needed to support human matchmakers and users without pretending attraction could be solved by a checklist. The product needed to remember what people said, notice what they responded to, and update the model after each interaction.
The preference gap
Users describe an ideal match in broad categories.
Real feedback arrives later, after a conversation or date.
Matchmakers carry context that rarely fits a standard form.
Simple filters miss tone, values, timing, and dealbreakers.
Why it matters broadly
Recruiting teams see this when a hiring manager rejects candidates who technically match the brief.
Real estate teams see it when buyers change priorities after touring homes.
Concierge and client-service teams see it when taste becomes clear only after the first recommendation misses.
The System
How feedback became product intelligence
We built Curate around a feedback-aware model. Instead of treating the initial profile as the final source of truth, the system learned from structured answers, conversational notes, visual preferences, and post-match outcomes.
The design principle was simple: make each interaction improve the next recommendation. That principle applies to any high-nuance product where user preference changes with experience.
Feedback dialogue
"I did not feel aligned with that person."
"What made it feel off: values, communication style, timing, or something else?"
"They were impressive, but they were not serious about the same kind of future."
The platform turned vague feedback into usable preference updates without forcing users to speak in technical categories.
Human review layer
Curate was not designed to make sensitive relationship decisions in isolation. It supported the people responsible for judgment, context, and client care.
The system made preference history easier to inspect and apply. The human team still owned the recommendation.
Technical Record
How the preference intelligence layer worked
The technical challenge was not simply ranking profiles. The system needed a way to combine different kinds of signals: structured profile fields, conversational feedback, visual preference patterns, and match outcomes.
We built a multimodal preference layer with vector search for similarity, natural language analysis for feedback, and a review loop that allowed each recommendation to teach the next one.
Structured Profile Context
Core criteria, dealbreakers, values, geography, timing, and stated preferences formed the first model layer.
Feedback Interpretation
Post-match reactions were converted into preference updates that could improve future recommendations.
Multimodal Signals
Visual and textual signals were handled separately, then brought together as decision support rather than a black-box verdict.
Recommendation Memory
The system preserved what had already been tried, why it missed, and what should change next.
async function updatePreferenceModel(matchId, feedback) {
const context = await extractPreferenceSignals(feedback);
await preferenceStore.update({
matchId,
confirmedSignals: context.confirmed,
changedSignals: context.changed,
nextRecommendationNotes: context.nextSteps,
});
return recommendationEngine.refreshCandidateSet(matchId);
}Citation
How should this case study be cited?
Suggested citation
Meru AI. "Preference Intelligence Case Study: learning from human feedback in high-nuance matching." Meru AI, updated June 2026. https://meruai.co/case-study/curate
What problem did Curate solve?
Curate helped a matchmaking platform learn from profile data, feedback, and match outcomes so recommendations could improve after each interaction.
Can this apply outside matchmaking?
Yes. The same pattern applies to recruiting, client matching, product recommendations, concierge services, education, real estate, and any decision where human preferences are complex and change over time.
Did the AI replace human judgment?
No. Curate supported human judgment by organizing preference history, surfacing relevant signals, and making feedback easier to apply.
Human Return
The human return
Better recommendations begin after the first miss.
The first stated preference is rarely the whole truth. The advantage comes from building a system that learns from every reaction, every correction, and every human decision that follows.