Loading
Designing probabilistic product discovery for 2,900+ retail items.
An AI-powered retail kiosk that helps shoppers describe what they want in everyday language and receive ranked, explainable product recommendations through conversational search.
In specialty retail, customers often walk in with vague needs, not exact product names. They may know the occasion, budget, taste, or person they are buying for, but not the right brand, category, or product vocabulary.
Traditional search and filters expect users to speak in catalog language. But shoppers usually speak in human language:
"I need a gift under $60."
"I want something smooth but not too strong."
"What goes well with grilled salmon?"
"I like smoky, but I'm still a beginner."
Glenen AI uses probabilistic search and RAG-based retrieval to map customer language to product metadata, rank relevant options, and explain why each recommendation fits.
Instead of returning a long list of products, the system gives shoppers a small set of explainable recommendations they can compare, refine, and act on.
Glenen AI was designed around one product belief: The best retail experience does not start with filters. It starts with understanding what the customer is trying to do.
In a specialty store, the most valuable employee is often the person who can translate vague customer language into confident product suggestions. Glenen AI brings that behavior into a kiosk experience: always available, fast, explainable, and connected to the live catalog.
Product vision: Glenen AI is not a better search interface for a liquor store. It is a decision-support system that helps every customer move from uncertainty to a confident product choice.
| Archetype | Situation | What they say | Design need |
|---|---|---|---|
| Occasion Buyer | Buying for dinner, party, or gift | "Just tell me what works." | Fast recommendation with a clear reason |
| First-Timer | New to the category | "I don't know the vocabulary." | Plain-language guidance without judgment |
| Flavor Chaser | Knows what they like | "I like smoky. What should I try next?" | Taste matching and intelligent exploration |
| Home Bartender | Has ingredients at home | "What can I make with what I have?" | Cocktail suggestions based on available inventory |
The hardest design problem in Glenen AI was not the screen layout. It was the dialogue. The assistant had to understand what the customer meant, ask the right follow-up questions, remember the conversation context, and move naturally between different shopping intents.
A shopper might start by asking for a gift, then switch to a food pairing, then ask for a cocktail idea. The interface needed to support that movement without making the user restart.
"I like something smooth, not too boozy, for a first whiskey."
Goal: Find something based on taste preference.
Approach: Map sensory language to product flavor metadata and return 3–5 matches with plain-language explanations.
"I have Aperol, gin, and lime. What can I make?"
Goal: Make a drink using available ingredients.
Approach: Cross-reference available ingredients with cocktail logic and recommend recipes based on what the shopper has or can buy in-store.
"I'm making grilled salmon."
Goal: Find a bottle for a meal.
Approach: Match the food profile to relevant product attributes and explain the recommendation in everyday language, not expert jargon.
"What's the difference between bourbon and scotch?"
Goal: Learn enough to make a better decision.
Approach: Teach through comparison, then pivot to a recommendation so education still supports product discovery.
The biggest product challenge was the vocabulary gap.
"Smooth, fruity, not too sweet, good for a gift."
Brand, category, ABV, price, region, tasting notes, shelf location, inventory status.
Glenen AI uses probabilistic search to translate messy customer language into ranked product matches. The system does not depend on exact keywords. It evaluates multiple signals and ranks products based on fit.
Smooth · Smoky · Fruity · Spicy
Gift · Dinner Party · BBQ · Cocktails
Beginner · Curious · Enthusiast · Expert
Red Meat · Seafood · Dessert · Cheese
Light · Medium · Full · Fizzy
Everyday · Special · Gift-Worthy · Splurge
Customer query: "I want something smooth for a first-timer who's into cocktails and has a $30 budget."
| Extracted signal | Mapped axis | Product logic |
|---|---|---|
| "smooth" | Taste Profile | Smooth, approachable, low-burn products |
| "first-timer" | Experience | Beginner-friendly, clear flavor notes |
| "into cocktails" | Occasion | Works well in mixed drinks |
| "$30 budget" | Budget Feel | Everyday or accessible price range |
The system retrieves candidates from the vector store, ranks them by multi-axis fit, and uses the AI layer to explain why each product matches the user's request.
Open-ended dialogue with intent classification.
Traditional search bar with AI-enhanced filters.
Filters require customers to know vocabulary they often do not have.
Every recommendation card includes a reason.
A plain list of product names and prices.
Customers buy confidence, not just products.
Return 3 to 5 ranked options.
One "best" recommendation.
A single AI answer can feel too final. Comparison gives users control.
Quick chips like "cheaper," "more premium," "more unique," "similar."
Restarting the conversation for every change.
Users should be able to steer the AI without losing progress.
Short responses, large tap targets, scannable cards.
Long chatbot-style responses.
A kiosk needs to be faster and more visual than a desktop chat.
Helps users start with guided prompts instead of a blank chat box.
Lets users type naturally or tap chips for budget, occasion, taste, or experience level.
Asks one useful follow-up only when it improves recommendation quality.
Shows ranked product cards with match reasons.
Helps users decide between similar products.
Lets users locate the item, save the result, or ask an associate.
The card does not just show a result. It explains the reasoning behind the result. That explanation is what makes the AI feel trustworthy.
Under $60 · Smooth profile · Beginner-friendly · Available today
Before: The kiosk opened with a blank input. Users had to figure out what to ask.
After: Added starter prompts like "Find a gift," "Shop by budget," "Pair with dinner," and "Compare products."
Before: The assistant returned recommendations as paragraphs. Text-heavy responses were hard to scan on a kiosk.
After: Created visual product cards with price, match reason, availability, and actions.
Before: The AI returned a single best recommendation. Users did not fully trust one answer.
After: Returned 3 to 5 ranked options, each with a different reason to choose it.
Before: Users saw what was recommended, but not why. The AI felt like a black box.
After: Added "Why this fits" labels and matched-signal explanations to each card.
The experience is supported by a modular AI architecture: a kiosk frontend, NLP backend, product data API, product catalog, and Night Agent QA workflow. Five components working together to deliver fast, explainable, on-catalog recommendations.
For inventory and product data, Glenen AI uses a human-in-the-loop process. AI-generated enrichments and catalog changes are not automatically pushed live. Operators review and approve changes before they reach the production catalog.
Why this matters for UX: This protects trust in a live retail environment. Customers should not receive recommendations based on unapproved or inaccurate product data. The diff-then-apply pipeline, enrichment proposals, and audit logs ensure operator approval before catalog changes go live.
3,264 SKUs in RAM — Fast access to product data during live kiosk use.
87% cache hit rate — Most lookups are served without slowing the conversation.
0% context bleed — One customer's conversation does not leak into another session.
+89% RAG quality improvement — Semantic retrieval outperformed the keyword baseline.
Faster product discovery — Users do not have to browse the full catalog manually.
Higher trust in recommendations — Each result includes a reason.
Better support for vague intent — Users can describe needs naturally.
More confident comparison — Users can compare ranked options without restarting.
Test whether users prefer one, two, or three clarifying questions before seeing recommendations.
Compare product cards with and without "Why this fits" explanations.
Test standing use, public typing comfort, visibility, and time pressure inside a real store.
Compare keyword search, filtered browsing, and probabilistic recommendation flows.
Glenen AI taught me that AI product design is not just about generating answers. The real challenge is designing the conditions around the answer: how users express intent, how the system clarifies ambiguity, how recommendations are explained, and how people stay in control.
"How can AI recommend products?"
"How can AI help shoppers feel confident about why a product fits?"
That shift shaped the product: guided prompts instead of blank search, recommendation cards instead of long responses, comparison instead of a single answer, and explanations instead of black-box AI.