Stable Profile
Proportions, facial structure, hair characteristics, comfort constraints and lifestyle requirements.
Product Thinking · Personalisation · Human-in-the-Loop AI
Beauty & Style Optimiser
Can highly subjective appearance decisions become a repeatable system by combining the person, the moment, the context and real-world feedback?
A personal experiment evolved into an explainable recommendation and learning system.
6personalisation layers
6decision dimensions
3recommendation modes
3real-world validation looks
01 / The distinction
“I like this” is not the same as “this works for me.”
A look can be attractive yet feel wrong because of proportion, maintenance, context or identity. I reframed personal style as a decision problem rather than a moodboard exercise.
02 / Core model
The best look is not static.
Stable self+Current state+Context+Available wardrobeRecommendationReal-world feedback→Model update
03 / Personalisation
Six layers organise the decision.
Visual Grammar
Personal rules about silhouette, contrast, balance, detail and identity—without reducing style to one label.
Current State
Skin and hair satisfaction, energy, confidence, visibility, effort and appetite for experimentation.
Context
Occasion, location, weather, formality, movement, photo importance and desired impression.
Wardrobe Data
Available pieces tagged by fit, visual weight, statement level, comfort, season and identity fit.
Outcome Feedback
Post-wear ratings compare the recommendation with lived experience and update the model.
How much complexity, experimentation or maintenance is appropriate for this person today?
04 / Decision Engine v0
One beauty score would hide the trade-offs.
Physical Fit
Does the look support the person’s proportions?
Identity Fit
Does it feel like the person—not merely look attractive?
Context Fit
Is it right for this place, occasion and desired impression?
Carry Capacity Fit
Is the complexity appropriate for today’s state?
Visual Coherence
Do the elements reinforce one another?
Effort Cost
What maintenance, styling and discomfort does it require?
Eliminate
Unavailable items, unsuitable dress code, weather conflict or a comfort limit that cannot be negotiated.
Influence
Visual clutter, weak waist definition or patterns that previously performed poorly—but may still work in a new combination.
05 / Recommendation design
Three useful answers—not one false certainty.
- ABest OverallStrongest balance across criteria.
- BMore ExpressiveHigher identity and visual impact.
- CLower EffortHigh comfort and Carry Capacity fit.
06 / Prototype
Fashion-led outside. Explicit logic inside.
The high-fidelity mobile concept uses a Modern Editorial direction: serif typography, neutral layouts, restrained colour and visible reasoning instead of dashboard complexity.
07 / Evidence confidence
Simulation proposes. Reality decides.
A visual or theoretical hypothesis.
Some evidence exists; more real-world use is needed.
A rule has performed consistently across repeated wear.
08 / First validation
Three worn looks refined the rules.
Contrast changed the rule
Romantic texture performed when balanced by sportier and more graphic elements.
A repeatable warm-weather formula
Structured open upper body + simple short + one waist detail supported the emerging visual grammar.
Low effort can still be intentional
A clean silhouette and defined upper body performed without high visual complexity.
Earlier assumptionRomantic styling does not work.
Refined ruleRomantic elements require contrast.
09 / Next phase
Test the full learning loop with users.
- Do explanations increase trust?
- Is daily-state personalisation useful?
- Are three differentiated recommendations better than one?
- Does post-wear feedback feel worth providing?
- Can repeated behaviour produce useful personal rules?