Product
Multi-agent AI platform
Web app
Desktop & Mobile
My role
Design lead
Head of Product
Team
3 Engineers
1 Designer
Timeline
April - Aug. MVP
Sep. - present V2.
%201.png)
ZAI is an AI-powered Web3 platform that transforms NFTs into intelligent agents, combining education, emotional connection, and investment insights to make Web3 more human and accessible.
Situation at the start: From chaos to clarity
When I joined the company, there was no product team, no roadmap, and no clear product direction — only six NFT characters and a vague idea of “combining NFTs and AI.”
The co-founders, both experienced in Web3, wanted to create something meaningful around our existing NFT collection but lacked a concrete vision of how AI could add value.
My first challenge was to turn this uncertainty into a tangible product strategy, balancing creativity, technical feasibility, and business value.
The NFT collection Narra released
RESEARCH APPROACH
RESEARCH APPROACH
Founder interviews & Internal brainstorms: Web 3 products common pain points
Sending a survey to know what users expect to get from us
I surveyed 244 users to understand what they value most when engaging with AI-powered characters.
The responses revealed four key user needs that shaped our product direction for ZAI.
PROBLEM REFRAMING
PROBLEM REFRAMING
Based on these insights, I reframed the initial problem from:
“How can we combine NFTs and AI?”
to a more actionable and user-centered challenge:
“How might we transform static NFTs into intelligent agents that engage users emotionally, educate them about Web3, and reveal real value opportunities?”
DEFINE
DEFINE
Early Direction Explored
Our unique opportunity was to blend emotion, education, and value discovery in a single system — something no one in Web3 was doing well.
Defining the Product Vision
Based on user insights and feasibility, I reframed the product direction around a simple but powerful idea, which This became the foundation of ZAI’s product vision — to make Web3 more approachable, human, and rewarding.
Transform NFTs into intelligent AI agents that can teach, connect, and empower users in the Web3 world.
Priorities and Roadmap
MVP
-
📝 Onboarding quiz
-
🕰️ Rewarding system
-
👬 Guest mode and Logged-in mode
-
🎮 Gamification system
-
💬 AI Chatbot
-
💸 Referral & Social Growth
-
🖥️ Desktop only
Future
-
📱 Mobile friendly
-
📰 News & Social content
-
🔍 Alpha mode
-
🤖 Advanced multi-agent Collaboration
Designing and Building the Multi-Agent Experience
Designing and Building the Multi-Agent Experience
From Vision to Execution
With the product vision in place, the next challenge was to translate our ideas into an engaging and technically feasible experience.
Our small cross-functional team — one AI engineer, one front-end, one back-end, and one designer — moved fast under tight constraints.
I managed the entire process using Notion for project tracking, monthly sprints, and twice-weekly product meetings to align priorities and decisions.
Product System/Logic Design
Partners & Guest Users: See an “egg” state, inspired by the Chinese term 崽
Narra NFT holders: See their assigned character’s full design and traits. We use this to reward NFT holders.
DESIGN
DESIGN AND PRODUCT LAYOUT
DESIGN
PRODUCT EXPLAINED
Guest mode - Lowering Entry Barriers for New Users
-
Users can start chatting immediately after onboarding.
-
Each guest session includes 10 free chats in Companion Mode and 3 chats in Coach Mode.
-
Once the limit is reached, users are prompted to create an account or connect their wallet to continue and unlock full personalization.
Login & Verification Flow
When users choose to log in, the system performs a quick verification to determine their access type:
-
Wallet: verify if there are supported NFTs in their wallets or they want to buy one
-
Email registration
Authenticated Mode — Unlocking Full Personalization and Progress
Once users decide to log in or connect their wallet, the system unlocks the full ZAI experience — linking emotional progress, learning data, and NFT identity into one continuous journey.
Onboarding quiz to collect user needs and preference
New user insturctions
After onboarding, users are welcomed by a short guided instruction — a tutorial that introduces how ZAI works and motivates users to start interacting immediately.，
Progression & Reward Loop — Building Long-Term Engagement Through Growth
To sustain engagement beyond the first interaction, we designed a progression system that blends emotional attachment with gamified learning. Every conversation contributes to the user’s ZAI Points, which represent both their emotional bond and learning progress.
As users accumulate up to 780 points, they gradually unlock traits, actions, and rewards that evolve their NFT agent visually and behaviorally.
Icons of the popup notifications for unlocking different traits, Generated by MidJourney,
More features
1. Change skin
Characters agent change skin and have new looks. For MVP, due to time limitations and complexibity, each character can only change colours. We used Midjourney to generate all skins.
Characters with skins in different colors
2. Daily check-in
AHA MOMENT!!
We designed character actions to surprise and delight users at just the right moments — what we call “Aha Moments.” These are subtle, emotionally resonant animations that are triggered by specific keywords or achievements during conversations.
Say “hi,” and your companion waves back.
Unlock a new trait, and they might throw a thumbs-up.
Say “thank you,” and you’ll see a shy, heartfelt reaction.
These aren’t random; they’re intentional design choices crafted to make the AI feel more alive, responsive, and emotionally engaging.
We don’t spoil all the fun — part of the magic is letting users discover these surprises organically, deepening their bond with each character over time.
Mobile design
AI Integration in Product
ZAI is not just powered by AI — it is designed around it.
Every layer of the experience — from conversation design to progression and scoring — is driven by intelligent systems that learn, adapt, and personalize over time.
1. Conversational Intelligence Layer 💬
Campanion mode - Chatgpt 4turbo
-
📍 GOAL
-
Create natural, emotionally engaging conversations.
-
-
❓WHY
-
offered strong contextual continuity, smooth tone variation, and superior ability to mimic human warmth and humor.
-
-
⬆OPTIMIZATION
-
Tuned prompt templates to emphasize empathy (“mirror user emotion before responding”).
-
Implemented short-term memory to reference recent dialogues, improving the sense of relationship continuity.
-
Added emotional feedback tagging (positive / neutral / negative) to fine-tune tone adjustments dynamically.
-
Coach mode - Gemini 1.5 flash
-
📍 GOAL
-
Deliver fast, structured, and trustworthy insights in finance and Web3 domains.
-
-
❓ WHY
-
selected for its high-speed retrieval, strong factual grounding, and lower latency — essential for information-heavy conversations.
-
-
⬆️ OPTIMIZATION
-
Custom instruction sets guide the model to format outputs into clear, digestible explanations.
-
Integrated lightweight fact-checking and citation modules (via internal APIs) for credibility.
-
Connected to real-time DeFi APIs and Web3 APIs, it provides contextual investment updates, NFT market data, and project insights.
-
2. AI-driven Scoring System 💯
Every user interaction in ZAI is evaluated by an AI-based scoring engine that measures the quality, tone, and impact of each chat session. The system goes beyond tracking frequency — it analyzes the depth and health of conversations. Each session produces a ZAI Point Score, which directly feeds into the evolution and reward system (unlocking traits, actions, and badges).
Key Scoring Dimensions:
-
Semantic Depth — evaluates message length, topic diversity, and contextual coherence.
-
Emotional Engagement — measures empathy, sentiment alignment, and conversational flow.
-
Token Consumption — quantifies effort and energy used within a session to balance system load.
3. Safety ⚠️
To ensure the emotional space remains positive and safe, we integrated a toxic language detection module into the scoring pipeline.
-
If the system detects toxic or abusive words (e.g., insults, hate speech, explicit content), the score is automatically reduced.
-
The user immediately receives a gentle warning message generated by the agent, maintaining the conversational tone but reinforcing safe boundaries.
-
Repeated offenses trigger temporary chat cooldowns to discourage misuse.
4. Feedback loop - Continuous Learning 👍 👎
The platform continuously improves through a closed AI feedback loop:
-
User messages are tagged by emotional intent and context.
-
Session quality metrics are fed back into the prompt library.
-
High-scoring dialogues are used to retrain the interaction model (few-shot tuning).
-
System parameters (temperature, memory weight, tone) are dynamically optimized per mode.
This cycle ensures the AI agent evolves along with user behavior — becoming more personal, engaging, and useful over time.
user testing
user testing
USABILITY TESTS BEFORE RELEASE
6 users in the community and 2 VCs
-
Bug Fixes with Direct UX Impact
-
Fix score not updating, Explain Emotional vs Professional modes clearly
-
-
UX improvements
-
Add animation when interacting with characters, clear instructions for the point system
-
-
Content and logic design clarification
-
Clear explanation about the score system and the revolution progress, visibility of badge wall
-
VC
Users
user testing
product feedback - auto workflow
I created an automated workflow using Zapier to sync product feedback from Discord directly into our Slack workspace. This eliminated manual monitoring across platforms and ensured the team never missed valuable user insights. It’s a simple example of how I extend my role beyond design—solving operational inefficiencies through automation to improve team responsiveness and focus.
Slack and Discord channels to directly collect user feedback to keep updating the product
OUTCOME
OUTCOME and next steps
Metrics
Feedback
1. Product
-
Current layout is more friendly for chatting, not friendly for long texts displaying
-
The majority of users missed the feature to switch modes (Companion vs. Coach)
-
Daily check-in is not attractive enough to increase stickiness
-
Navigation Optimization
-
Clear guidence, especially for new users
2. Chatbot
-
Memory: the chatbot can not remember the past context
-
In coach mode, the answers are too general, not providing detailed content with depth
-
Looking for more inforamtion eg. news, trends, markets, etc.
improvement
V2. improvements
Campanion mode chatting page update
Campanion mode Before vs. After
Improve the Coach Mode experience, auto-provide topics to users to aviod friction to make it easier for users to get started
Exploration 1
Exploration 2 ✅
Navigation update
Onboarding quiz update
Based on the insights gathered from our first onboarding survey — which revealed high user interest in simulated trading, emotional support, and beginner-friendly Web3 content — I created an updated version of the questionnaire to capture more detailed user needs and preferences. This second version includes more nuanced questions about Web3 experience, desired AI interactions, and specific investment goals.
We also give 10 points as rewards for completing the quiz.
👉 View the updated survey here
Evolving ZAI’s AI Capabilities
Next step
1. Context Memory & Long-term Learning
-
Problem:
-
Current AI agents reset after each session, limiting emotional continuity.
-
-
Next Step:
-
Implement contextual memory that allows ZAI Agents to recall past interactions, track user preferences, and evolve their personality over time.
-
2. Coach mode capacity improvement
-
More APIs and data source
-
Improve interaction and expand the data type that can be provided
3. Real time Web 3 intelligence
-
Real-time Market & DeFi Data (API Integration)
-
The AI agent fetches live data from APIs, filters results, and summarizes findings in natural language with visuals.
4. Improve product discoverability
-
Build SEO-optimized content around AI x Web3 education topics to capture search traffic.
-
Partner with influencers or NFT communities to drive external awareness.
-
Add shareable moments (e.g., “Your ZAI just evolved!”) to organically drive visibility.
