Reported average chat duration
MEL · CONVERSATIONAL AI
mel
Designing the path from a first chat to deeper use.
We first suspected short sessions were an AI-quality problem. User journeys showed something else: the capability already existed, but people could not see what to try next. I brought relevant modes into the conversation and reduced complex choices into clearer starting points.
LISTENINGTHINKINGSPEAKING
02PRODUCT OVERVIEW
A LIVE AI CHAT EXPERIENCE
Conversation, character, and space respond as one experience.
MEL is a real-time AI character experience. People can talk by text or voice, create their own character, and enter visually generated contexts that make the relationship feel more present.
- REAL-TIME CHAT
- VOICE
- CHARACTER CREATION
- GENERATED CONTEXTS
03PRODUCT INSIGHT
WHAT THE JOURNEY REVEALED
Users did not need more features. They needed a path to the next one.
The issue was not AI quality alone. Advanced modes stayed invisible until the interface introduced the next useful possibility at the moment it became relevant.
04CHALLENGE 01 · VOICE CONVERSATION UX
TURN-TAKING CLARITY
Voice made the AI feel alive. State clarity made it usable.
A live exchange can become ambiguous quickly. The interface had to make speaking, listening, waiting, and response timing legible without interrupting the conversation.
- 01Problem
Voice, text, and visual response competed for attention.
- 02Principle
One dominant state at a time; transition before decoration.
- 03Outcome
Speaking, listening, waiting, and responding became distinct states.
05CHALLENGE 02 · CREATE YOUR OWN MEL
RELATIONSHIP ONBOARDING
Creation needed freedom, but also a place to begin.
Character creation was not merely avatar setup. It was the first emotional commitment to a relationship. I grouped the choices into clearer starting routes, preserving expression without presenting an infinite wall of options.
- 01Problem
Too much freedom created hesitation before creation began.
- 02Principle
Simplify the first choice; reveal expression progressively.
- 03Outcome
A clearer route from intention to character to conversation.
06CHALLENGE 03 · GENERATIVE DATE SCENE WORKFLOW
DESIGN × ML COLLABORATION
Generated visuals had to feel usable, not merely possible.
In late 2023 and early 2024, consistent people and backgrounds were difficult to generate. I worked with the ML team through a ComfyUI-based workflow, iterating on mood, composition, continuity, and scenario clarity until date scenes could support the product experience.
- 01Brief
Mood, relationship context, place, and product constraints.
- 02Generate
ComfyUI workflow with repeatable inputs and controls.
- 03Refine
Character consistency, composition, artifacts, and legibility.
- 04Ship
Date and place imagery selected for real product scenarios.
07SYSTEM & COLLABORATION
ONE INTERACTION LANGUAGE
Give multiple modes one shared interaction language.
Voice, text, creation, and visual states needed to feel like one product across iOS, Android, and Web. Shared state patterns and naming reduced reinterpretation during design and engineering handoff.
- 01Before
Modes and platforms expressed the same states differently.
- 02System
Shared states, component structure, naming, and behavior rules.
- 03Effect
Clearer handoff and a more extensible multi-platform product.
08IMPACT & REFLECTION
REPORTED BETA SIGNALS
Capability became meaningful when people could enter, understand, and continue.
Reported onboarding-exit change
Beta participants
Values are retained from the original beta materials. Where the exact definition or experimental method is incomplete, they are presented as reported directional signals—not causal proof.
Emerging AI products do not become meaningful through capability alone. They become meaningful when users can enter, understand, and continue the experience.
From guidance to sensory feedback
After reducing hesitation in an AI creation flow, I explored a different question: can immediate sound, motion, and haptics make interaction itself worth repeating?