Product case study
Frolio
A fragrance collection that helps you choose.
What Frolio is
Frolio is a website for people who collect perfume or want to find their next fragrance. You can organize bottles you own, samples you’ve tried, and a wishlist, then keep a diary of what you wear, your ratings, and private notes. You can also search the fragrance catalog in your own words, find suggestions based on familiar smells, or return to daily recommendations informed by your collection.
What I built
I own Frolio’s product strategy and build it solo with Claude Code. I centered the product on a collection and wear journal, then connected three discovery paths to different needs: getting started, searching for a scent, and choosing what to try next. The catalog contains 150,000 unique fragrances.

A collection is more than a list.
What someone owns, has sampled, and actually wears are different signals. I made those distinctions the foundation of the product: preserve a useful personal record, then use it to support discovery and buying decisions. Beginners can start with familiar smells instead of expert scent vocabulary.
Start, collect, wear, discover.
I prioritized the collection and journal as the core experience, with discovery and buying tools connected to that record. Product choices keep public sharing separate from private notes and require people to review photo-recognition suggestions before they enter the collection.
- Collection and privacy. Owned, sampled, wishlist, and archived states organize bottles. A public collection can include ratings and recent wear history; personal notes remain private.
- A wear journal and a habit. Wear logs, ratings, a calendar, private notes, and logging streaks help people revisit what they actually use. Weekly tasks and a capped freeze balance support the streak.
- Buying with context. Researched US retailer offers identify bottle size, concentration, source, and checked date. Before you buy compares listed notes and accords with owned bottles and samples.
- Reviewable photo intake. A bottle or shelf photo can start collection entry; each proposed bottle is reviewed by the user before it is added.
- Three interface languages. English, Simplified Chinese, and Traditional Chinese interfaces share the product flows, with reviewed scent terminology.
Three ways into discovery.
I separated discovery into three tasks, with an interaction and ranking method suited to each. Guided matching gives beginners a structured start; semantic search interprets open-ended descriptions; daily picks use collection history and feedback. Rules handle matching and daily ranking, while the LLM supports interpretation where language is less predictable.
- A familiar-smell taste finder. A rule-based taste finder returns up to three catalog matches without a model request. Three saved learning quests help beginners take the next step.
- Semantic search and similar fragrances. Claude interprets a scent description; Voyage AI produces 1,024-dimensional embeddings. pgvector and an HNSW index retrieve candidates for reranking and the similar-fragrances rail.
- Personal daily picks. Up to three private daily discoveries use collection ratings, wear history, scent facets, and negative feedback. Deterministic ranking adds variety and duplicate safeguards without a model call.
Frolio architecture
Collection states, ratings, wear history, and private notes live in PostgreSQL through Prisma. Daily picks use deterministic ranking; semantic search uses Claude query interpretation, Voyage AI embeddings, and pgvector retrieval. Next.js and NextAuth provide the interface and account boundary, while Docker and Cloudflare serve the product.
The verified components behind Frolio’s public product.
Product experience, discovery, and production infrastructure.
A product that has to keep running.

Running the product is part of building it. Frolio runs in Docker behind a Cloudflare tunnel, with rate limits, error monitoring, content moderation, and scheduled backups.
- Application safeguards. Upstash handles rate limits, Sentry monitors application errors, and Anthropic supports safety moderation.
- Scheduled backups. A systemd timer runs scheduled backup jobs.
Current numbers
Catalog scale as of October 27, 2026, confirmed by me as the product owner.
Unique fragrances in the catalog. This measures catalog scale; it does not measure adoption or recommendation quality.
Guardrails
These guardrails help users check suggestions before acting on them: fragrance comparisons are clues, prices include source and date context, and photo matches need confirmation.
- Note and accord overlap is a comparison clue, not proof that two fragrances smell alike.
- Retailer offers cover researched US bottle variants; dated observations and Fragrantica shop ranges are labeled separately.
- Photo recognition requires user review. Forum image uploads on Cloudflare R2 remain disabled.







