AI-first perfume search for people who describe scent through moods, memories and voice.
Semantic search disguised as a shop
Scentum is an AI-first search layer for fragrance. Voice prompts and natural-language queries are translated into vector weights, vibe clusters and ranked matches — turning notes, accords, and ratings into a navigable semantic space.
- Voice-to-intent search for full-sentence prompts
- GenAI maps mood, memory and use case into retrieval weights
- 5D ScoreVector for similarity and recommendation space
- Vibe tags act as cultural embedding vocabulary
- Ranked results explain why each scent matches
Vector space before product grid
A single 1-to-100 rating collapses fragrance into a hierarchy. Scentum maps every perfume into five orthogonal dimensions so GenAI can turn language like "warm, skin-close, office-safe, a little strange" into a searchable position in vector space.
- Authenticity ratio of natural to synthetic feel
- Projection from skin-scent intimate to room-filling loud
- Longevity hours on skin before it's gone
- Complexity how much it evolves between top and base
- Versatility how many moods and seasons it survives
Three pillars
The product is built from three AI-search systems: a vector model, a vibe taxonomy, and a voice-friendly GenAI discovery flow that grows on top of both.
- Five orthogonal dimensions per perfume
- Derived heuristically from notes & ratings
- Refined offline by Claude Haiku enrichment
- Drives filters, similarity, and recommendations
- Visualised as radar, bars, or interpretive prose
The substrate everything else stands on. Five clean dimensions that voice and conversation can modify fluidly.
- Cultural & aesthetic clustering
- Hand-picked vocabulary, not auto-generated
- Maps perfumes to lived experience, not chemistry
- Multi-tag with AND / OR logic
- The bridge between data and feeling
Where the catalog becomes a language. quiet-luxury, archive-core, late-french-theory — each tag is a tribe and a recommendation engine in three syllables.
- Faceted filters as the base layer
- Natural-language search via Claude
- Voice input for full-sentence intent
- Semantic retrieval across vector space
- Conversational refinement loop
- Personalised "Scent Story" for each match
The AI is the search interface: it parses intent, moves through vector space, explains matches and narrows the possibility space through dialogue.
Vibe Tags
Concrete examples of the cultural vocabulary the catalog speaks. Each tag is a cluster of perfumes that share a feeling, not a formula.
The AI-first search interface
Familiar e-commerce shell, semantic-search soul. Voice maps to intent; sliders map to dimensions; chips map to vibes; every card explains a multidimensional match at a glance.
Under the hood
A small, dependable AI-search pipeline. Public web data, offline GenAI enrichment, vector-ready catalog metadata and a frontend that treats the database as a semantic space, not a stockroom.
Scrapy + Zyte
Two-phase spider on Zyte Cloud crawls multiple fragrance sites, deduplicating across runs to surface only newly published listings.
Claude Enrichment
Haiku derives the missing fields — authenticity, complexity, versatility, vibe tags — from notes, accords, and reviewer language.
MongoDB Catalog
Document store holding each perfume with its full ScoreVector and vector-ready metadata for faceted queries, semantic retrieval and similarity ranking.
React / TypeScript
Mobile-first interface with voice input, semantic sliders, vibe chips, AI-first search and an interpretive product page.
From search to conversation to guidance
The four-step shift from keyword search to voice-led semantic guidance. Each step is independently shippable; together they compose a fragrance consultant, not a search box.
Voice as Input
Speech-to-text feeds raw utterances. Claude maps them to the ScoreVector and explains the mapping — interactive, not opaque.
Fragrance Oracle
An optional 3-to-5 question guided flow. Each answer animates a live ScoreVector radar — results feel earned, not browsed.
Refinement Loop
"Too sweet", "warmer", "less woody" — quick reactions adjust weights live. Preference learning inside a single session.
Scent Story
The product page becomes a mirror — fragrance described back through the user's own language and revealed memories.
The interpretive layer is live
Scentum is in active development at scentum.space. The catalog grows every scrape; the AI-first search layer ships in increments across voice input, semantic retrieval and vector-space recommendations.