Project Goal: Building AI SaaS from Scratch in a Hardware Company
As a Solo Product Designer, I led the 0-to-1 evolution of ChirpAI, ViewSonic’s first AI-native SaaS. This wasn’t just a design project; it was a digital transformation effort to bridge a hardware-centric legacy with the future of AI-driven classroom intelligence. I orchestrated the entire lifecycle—from decoding fragmented market policies to delivering a pilot-ready MVP.
Key Achievement: Successfully achieving a 100% pilot intent from US school administrators by solving a critical data-visibility problem.
The Problem: Breaking the “One-Size-Fits-All” Barrier
A Silent Crisis in US K-12 Classrooms
In a typical US classroom of 30 students, teaching is often a "One-to-Many Broadcast." While the teacher delivers a lesson, the "silent majority" of students often fall through the cracks because their individual needs are invisible in real-time.
Quantifying the Gap
- Teacher Burnout: 33% of educators are too overwhelmed by administrative tasks to provide 1-on-1 mentorship.
- Student Isolation: 19% of students navigate their entire learning journey in total silence, receiving zero personalized guidance.
- The Trust Barrier: 25% of students feel no meaningful connection with any adult at school, making them less likely to ask for help.
First Concept: Broadcast the Lesson. Give Every Student Their Own AI Agent.
In a traditional formative assessment, a teacher asks a question, the class answers, and the teacher gets back an average score — a view of the whole, but not each individual. Or they walk the room, collecting feedback one student at a time.
Our hypothesis: use AI to scale what was once only possible 1-on-1. A teacher broadcasts one prompt to the entire class; every student responds to their own AI agent; the AI automatically groups students by proficiency and returns a prioritized list of who needs intervention — expanding the time-intensive 1-on-1 dynamic into a 1-to-30 classroom reality, powered by differentiated learning at scale.
| Phase | AI-mediated Logic | Educational Value |
|---|---|---|
| 1 Assessment Creation | Curriculum-Led Generation: Automated assessment creation based on selected teaching standards. | Removes the manual “prompt engineering” burden for teachers. |
| 2 Student Interaction | Guardrail-First Chatbot: A “Topic-and-Level-Constrained” interface for student engagement. | Ensures pedagogical safety and prevents hallucinations. |
| 3 Insight and Intervention | 3x3 Matrix Analysis: Automated clustering into 9 proficiency tiers for a Second-Round Intervention. | Scales Differentiated Learning without increasing teacher workload. |
Failing Fast: Three Iterations Toward the Right Direction
Through early experimentation, we identified critical gaps in our initial direction. These iterations allowed us to refine our product-market fit before scaling.
| Iteration | Hypothesis | The Gap | Strategic Pivot |
|---|---|---|---|
| Trial 1 | Post-Class Assessment: A quiz tool to verify understanding after class. | High Friction: Students felt over-tested; teachers felt over-burdened. Low participation due to lack of instant feedback. | Shift from “Post-class Testing” to “In-class Interaction” to capture real-time engagement. |
| Trial 2 | Topic-Based Insights: A dashboard to cluster performance by topic. | Low Actionability: Required student-platform ecosystem support that was outside the product scope. | Shift from “Data Visualization” to “Actionable Pedagogy” using LLMs to group students for immediate intervention. |
| Trial 3 | 3x3 Granular Matrix: A 9-grid system (High/Mid/Low) to categorize students. | Signal vs. Noise: 1. AI Limits: Dialogue data is too unstructured for accurate 3-tier grading. | Optimized to 2 levels: Priority Topics x Binary Tiers (Emergent Understanding x Deep Comprehension) to ensure accurate, high-stakes intervention. |
From Internal Insights to Market Strategy: Shaping the MVP Direction
The initial pivots were more than internal trial and error. Before committing to the final MVP development, we summarized the internal vision and external market constraints, ensuring we focused and navigated to the right problem with the right technology.
Constraints
- Scenario Fragmentation: Confusion between pre-class, in-class, and after-class user scenarios.
- Market Saturation: A crowded landscape of established AI "test" and "practice" tools.
- Data Scalability: Extracting stable checkpoints is hindered by fragmented regional curricula, requiring extensive timelines for data crawling and RAG architecture development.
Enablers
- Policy & Budget Shifts: Favorable changes in US government educational funding.
- Refining "Insights": The demand for data-driven insights remains strong, provided the content is more actionable and integrated.
User Research: Anchoring the MVP in Real Market Signals
To navigate the ambiguity of a new market, I spearheaded a series of Lean Discovery cycles. Rather than pursuing formal academic rigor, we focused on rapidly triangulating high-impact pain points from fragmented market signals. This agile approach allowed us to move fast while ensuring the 0-1 trajectory was anchored in objective US policy and stakeholder needs.
| Studies | Methodology | Finding | Design Impact |
|---|---|---|---|
| Policy & Ecosystem Mapping | Desktop Research: Synthesized complex US frameworks (MTSS, ESSA, SEL) and intervention systems. | Counseling interventions require close expert involvement and are not suitable for productization. | Dropped counseling-focused value proposition and budget allocation. |
| In-Class Activity Habits | Interviews & Routine Field Observations: Mapped activity workflows with teachers and US trainers to identify pain points. | Discovered “Exit Ticket” as an established in-class teaching habit. | Feature Integration: Introduced “Exit Ticket” features aligned with curriculum standards. |
| B2B Stakeholder Mapping | Stakeholder Interviews: Analyzed AI perceptions and concerns within the “Class-School-District” hierarchy. | Understood the gap between users (teachers) and buyers (districts), particularly around procurement policy constraints. | Separated buyer and user considerations; addressed procurement compliance as a core product requirement. |
| Cross-Cultural Concept Testing | Comparative Concept Testing: Evaluated 3 prototype variations to probe instructional mental models and expectations. | Taiwanese teachers prioritized extended practice materials; US teachers valued student-led expression and flexible grouping strategies. | **Shifted AI usage **from “differentiated practice” to “student empowerment” to fit the US market context. |
Information Architecture: The “Talk”-Centric Model
Constructed a “one-to-many” interaction model using “Talk” as the foundational unit. By referencing the state management of survey platforms, I designed a scalable flow to handle asynchronous interactions between teachers, students, and AI. This core logic unified the entire system, from sitemaps and permissions to the final UI layout.
Design Highlights
Create: Scenario-Driven Talk Creation
- Streamlined the AI setup process by offering context-specific templates based on core classroom scenarios. This “template-first” approach minimizes decision friction and ensures each session is powered by a tailored AI prompt structure, input logic, and insight goals.
- Provided language options for non-native speakers (e.g., Spanish-speaking or immigrant students) and bilingual programs.
Chatbot Configuration
- AI-Assisted Metadata: The system automatically extracts educational standards and generates chatbot descriptions from uploaded content.
- Live Preview: Integrated a real-time testing panel for instant verification of the AI’s persona and logic before going live.
- Asynchronous Updates: To avoid blocking teachers during iterative adjustments, the UI is never fully locked — individual read/write constraints from database updates are reflected directly in the interface.
Inclusive & Safe Student Interaction
- Bilingual Support: A “Swift Switch” language toggle allows students to interact in their preferred language.
- Real-time Safeguards: Integrated an AI-driven monitoring system that automatically detects 18 sensitive topics, triggering immediate notifications for teacher intervention.
- Reply Prompts: Guided response options help students who struggle with expression or typing to quickly convey their current understanding.
Real-time Talk Results: Teachers can capture student thinking without closing the session.
- Enabled teachers to capture the “pulse” of the classroom without ending the session. The dashboard provides a Live Activity Feed for instant feedback, while AI automatically summarizes safeguard triggers and extracts key student quotes to highlight learning gaps.
- Provides per-student suggestions on misconceptions and topics of interest.
UX Writing Guide for Education
To ensure a consistent and expressive product voice, I established a UX writing framework based on NN/g guidelines, co-developed with a U.S. Education Ph.D. This guide serves as the foundation for our AI’s personality and safety:
- Pedagogical Tone Strategy: Strategic use of encouraging vs. authoritative tones to guide students effectively.
- User-Specific Adaptation: Tailored vocabulary and tone nuances for distinct teacher and student interactions.
- Interaction Safeguards: Defined communication boundaries and “taboos” to ensure AI-user safety and professionalism.
- Linguistic Standards: A centralized system of Voice & Tone adjectives and a product-wide Glossary.
Product Market Fit Evaluation
Hypothesis
- Addressing Differentiated Learning Needs: “In the past, my teaching methods couldn’t effectively accommodate students who were falling behind. ChirpAI provides an excellent channel to bridge that gap.”
- Empowering Student-Centered Teaching to Boost Motivation: “The chatbot guides me on how to express my thoughts, which makes me feel truly understood.”
Recruiting and Study
US (Target Market):
- Participants were recruited through online seminars, on-site teacher training sessions, product demos, and the ViewSonic US educator community.
- From an initial pool of 49 participants, we collected 10 valid samples (3 online and 7 on-site), including both IT specialists and educators.
Taiwan (Pilot Market):
- Focused on elite schools and international bilingual schools adopting American-style education, recruited through on-site visits and online events.
User Validation Results (3 Online + 7 On-site)
Through structured interviews and concept testing, we quantified the product’s impact across three key dimensions:
- 10/10 Pilot Intent: Secured trial commitments from all interviewed US school administrators.
- 6/7 Purchase Advocacy: High intent from trial educators to advocate for institutional procurement.
- 9/10 Efficiency Gain: Users reported a significant reduction in feedback loop and administrative overhead.
The Core Differentiator
The primary driver behind these results is ChirpAI’s distinction from ChatGPT:
Unlike a standard chatbot, this platform empowers teachers to manage and monitor AI-student interactions in real-time—fostering a safe, pedagogically-sound environment that motivates learning.
Interview Quotes
- Existing ViewSonic customers value the potential for student data integration and future cohesive solutions.
- Non-ViewSonic customers see ViewSonic’s AI innovation as a strong branding opportunity.
Market Traction: Taiwan Pilot
Cold Outreach: The product team sourced potential partners through cold emails and school referrals.
Product Adoption: After on-site demos, one school had teachers adopt the product into lessons the very next week; another inquired about large-scale teacher training — demonstrating low adoption cost and strong product-market fit.
Scalable Monetization: Helped the team secure a district deal covering 3,000 students, proving the design's scalable commercial value.
Reflections: My Framework for AI-mediated Interaction
Building ChirpAI taught me that AI success is measured by the quality of human augmentation. I’ve distilled my learning into four pillars:
- Intelligence: Transforming raw data into context-aware insights. True value is created when the user perceives “intelligence” through actionable classroom-bound suggestions.
- Modality: The choice of interaction (chatbot vs. file-input) is a strategic one. It must be dictated by domain-specific scenarios, not default patterns.
- Human-in-the-loop: It’s about the timing of AI intervention. The goal is to substitute manual effort while preserving creative space for user customization.
- AI Ethics: Defining the limit of AI commitment. Trust is built by preventing over-reliance and keeping the AI strictly within pedagogical boundaries.
Final Acknowledgments
This 0-to-1 journey was a rare and rewarding challenge. I am grateful for the opportunity to have shaped the entire user experience from the research phase onward. Deep thanks to my PM, Abbie, for the late-night brainstorming and discussion.
