Wardrobe Ingestion
- Normalize wardrobe images
- Extract type, color, silhouette
- Store embeddings for matching
Duorin https://www.duorin.com
An AI styling assistant that turns wardrobe, weather, calendar, and preference signals into daily outfit decisions users can trust.
Why this matters
Personal styling is a high-frequency decision surface: if users do not trust the product, they churn before model quality gets a fair test.
The hard problem is not generating attractive outfits. It is earning trust through explainable, contextual recommendations that hold up on real mornings.
The product reduces decision fatigue through clear logic and actionable outcomes.
The Problem
Most styling apps fail because they optimize for option volume, not decision quality. Recommendations look good but ignore context, so users cannot trust them when they need to get dressed.
- Trust breaks when users cannot follow why an outfit was selected.
- Calendar, weather, and wardrobe signals must compound in a clear order.
- Too many weak options increase cognitive load instead of reducing it.
Core Insight
Styling is not an inspiration problem — it is a decision problem.
Most styling apps optimize for inspiration. Duorin optimizes for decision certainty.
Product Evolution
Before
Item-level verdict
Users evaluated one garment at a time. Outfit assembly still happened manually.
After
Clearer item-level decisions
New garments are evaluated before they enter the wardrobe. Decisions are clearer and easier to act on.
This improved item-level clarity, but users still had to assemble outfits themselves.
Duorin separates two decisions: what enters the wardrobe, and what gets worn. Clear verdicts help users decide what enters the wardrobe, reducing noise in later outfit decisions.
System Overview
Duorin is a decision pipeline, not a single feature.
Wardrobe Upload
Item Evaluation
Outfit Generation
Item Evaluation / Style Scan
- Score compatibility with wardrobe
- Check undertone and preference fit
- Detect overlaps and gaps
- Filter weak items before entry
Outfit Generation
- Apply constraints in sequence
- Calendar → weather → wardrobe
- Generate a small set of viable outfits
Better inputs → better decisions → better recommendations
Controlling what enters the wardrobe improves the quality of downstream outfit decisions.
Early Signals from Beta
Duorin is currently in a live beta with around 20 active users.
These are early directional signals, not statistical conclusions.
- Users hesitated when recommendation reasoning was not explicit
- Item-level clarity improved, but outfit decision fatigue remained
- Wardrobe quality strongly affected recommendation quality
A consistent pattern emerged:
Improving input quality and decision clarity upstream has an outsized impact on trust and usability.
Key product decisions
- Context before inspiration
- Explainability before magic
- Fewer, stronger recommendations over endless browsing
- Honest empty states over weak AI guesses
Design Approach
- 1.
Design for daily decision-making, not browsing.
Why it mattered: Repeat use depends on reliable outcomes during ordinary days.
- 2.
Make outfit logic explainable at the point of decision.
Why it mattered: Trust increases when users can see why each recommendation appears.
- 3.
Handle missing wardrobe data and weak matches as explicit states.
Why it mattered: Clear gaps preserve trust better than vague recommendations.
Outfit Generation Logic
InputCalendar events, weather conditions, available wardrobe items, and user style preferences.
LogicApply constraints in sequence, filter invalid items, score viable combinations, and prioritize diverse outfits that fit the day.
OutputA small set of context-ranked outfit decisions with visible reasoning and clear next actions.
Impact
- Recommendations became predictable instead of arbitrary.
- Users could trace why each outfit appeared.
- Decision time reduced to a few viable choices.
Reflection
Current limitation:
The system initially emphasized output quality over decision transparency. Users could not trace the reasoning behind recommendations, which reduced trust even when results were correct.
Next iteration focus:
Make reasoning first-class in the UI, expose why each recommendation appears, add confidence indicators, and support quick item-level adjustments to keep users within the decision flow.