Sensory profiles
Up to three taste identities — e.g. Gas / Fuel bombs, Evening knock-out, Sweet tropics — each defined across aroma, flavour, effect, texture, potency and priority weights. A completeness ring unlocks matching at 60%.
Sensory sommelier for cannabis
Describe how it should smell, taste and feel — find the match on any menu.
A private-by-design tasting room. Instead of “indica / sativa / hybrid” reductionism, SŌMA lets a member build a personal sensory profile — aromas, flavours, effects — then scans any dispensary menu and ranks every strain against that taste with a deterministic, fully explainable match engine.
The idea
A private tasting room, not a smoke-shop menu.
Every dispensary menu sorts flower by legacy taxonomy (indica / sativa / hybrid) and marketing copy (“euphoric, uplifting, creative”) — categories that don’t actually describe what a person’s palate wants tonight.
SŌMA borrows its language from wine and perfumery instead. Members describe how they want a flower to smell, taste and feel, and the app finds the ones that match — on any menu, with a written verdict for every result. Discover, experience, evolve.
What it does
Eight rooms in the tasting house.
Taste Match Engine
A deterministic scoring engine reading the same sensory vocabulary as the questionnaire, catalog and audit log. No black-box ML — every match ships a per-sense breakdown and a written verdict.
Menu analysis
Paste a dispensary menu as text — a deterministic parser pulls out grower, THC, price and weight — and get a ranked #1 / #2 / #3 board plus the full list, each row expandable into the analysis, with a compare-two side-by-side view.
Taste Blender
A virtual fourth profile that mixes two or three real ones with a bias slider and a bridge mode (min-across-worlds, so only strains strong in every world survive). Solves “I want both moods tonight.”
Harvest
An editorial catalog of 895 curated strains — sommelier-style curator notes, lineage, market aliases and a sensory family taxonomy (all in the language the engine reads from), with hand-drawn poster art for the headline strains.
Collection & Compare
A member’s shelf of tried and wishlisted flowers, plus a two-strain compare view that decomposes match scores by profile and by sense.
Two ways to arrive
A member can answer the questionnaire — or simply describe a flower they loved, or an experience they had, and let the app infer the profile from it. The inference is shown back, with its contradictions named, before it becomes a taste identity.
Learn
Eight editorial pieces on how to read a flower — terpenes, sensory families, why potency is the least interesting number on a label — in the same sommelier voice as the verdicts.
Zkittlez × Gelato
91%“Candied tropical nose over a creamy gelato base; a soft, rounded evening lean. A near-perfect fit for your Sweet tropics profile.”
Under the hood
Selected engineering problems.
Deterministic, explainable scoring
Engine, questionnaire and strain data all draw from one canonical
vocabulary object (VOCAB_VERSION), so a score is always
reproducible from stored inputs — audit rows survive vocab migrations
by pivoting on the version.
Anonymous → registered, no migration
Every visitor gets a signed anonymous UUID cookie; profile, history and audit log accumulate under it. Registering just sets username, email and password hash on the same row, so nothing is lost. HMAC session tokens validate in-process so a DB blip never silently logs a member out. Verification and reset tokens are stored only as hashes, single-use and expiring; deleting an account removes the row.
A gate before the door
The whole site sits behind a 21+ acknowledgement that states, before anything else, what SŌMA is and isn’t: an educational sensory resource, not a seller, not affiliated with any dispensary, and it never tells anyone where to buy. The acknowledgement is versioned, so the gate returns if the disclaimer’s substance changes.
895-strain catalog performance
First paint started in “did it freeze?” territory. Fix: cache the assembled catalog (hundreds of thousands of comparisons no longer paid per cold start), trim the per-card payload sharply, paginate to 40 cards with IntersectionObserver infinite scroll, and add an instant skeleton.
Champagne-gold, built in CSS
Every gold surface is pure CSS — no image assets: masked hollow-ring borders, sculpted medallion rings with a debossed cream well, star-shine glints, and a 42% frosted card that lets the embossed leaf backdrop show through. All theme-aware.
Design language
Gilded, editorial, sommelier.
Gilded champagne-gold and frosted-cream cards on an embossed cannabis-leaf backdrop. Display in Fraunces (a high-contrast antiqua), UI in Inter. Every microcopy line reads like a sommelier note, not a marketing bullet — sensory-family gradients shift per strain territory (gas-og, garlic-funk, citrus-haze…).
Cream · #F6F3EE
Champagne · #C8A76A
Shine · #F6D88A
- Fraunces + Inter
- Frosted-cream cards
- Embossed leaf backdrop
- Per-strain poster art
- Curator notes on 895 strains
The build
Stack.
Next.js 15 (App Router · React 19 · Server Components) with TypeScript and Tailwind. Prisma over PostgreSQL (Supabase); custom scrypt + HMAC-cookie auth (no auth library). OpenAI is optional — it only rewrites the written verdicts and infers off-catalog strain names; it never touches the scores. Tested with Node’s built-in test runner plus a stress harness that guards against “the illusion of intelligence.”
Next.js 15React 19TypeScript
Tailwind CSSPrismaPostgreSQL · Supabase
OpenAI (optional)node:testlucide-react
Status
Functional MVP · pre real-menu validation
The tasting room works. The menus haven’t been walked.
Profiles, the match engine, the Blender, Harvest, Collection, Compare, accounts and the audit log all run today, against a catalog of 895 curated strains. What is not settled:
- No validation against real dispensary menus at scale. The parser reads pasted text well; how it holds up across many real menus is untested.
- No payments and no membership tier. Everything is open.
- Curator notes are ours, not a panel’s. They are written from published sensory descriptions, not from a blind tasting — and the app says a match is a sensory prediction, never a guarantee.
- Real quality still depends on grower, freshness, packaging date and storage, which no engine can read off a menu.
◆ ◆ ◆
Discover. Experience. Evolve.
SŌMA is live — step into the tasting room.