B2B
AI Interaction Design
LLM System Instruction
AI Assistant
Multi-Agent
Overview
Flow is an ADHD-friendly AI assistant and toolbox built with Microsoft Azure AI Foundry. It helps neurodivergent professionals transform challenges into opportunities by amplifying their unique strengths. Through behavioral research, system prompt design, AI interaction patterns, and evaluation methods, Flow demonstrates how inclusive AI enables professionals to thrive in their workflows.
Corporate Sponsor
Timeline
9 months (Dec 2024-Aug 2025)
Microsoft Team
Principal PM, Azure OpenAI (Sponsor)
Senior UX Researcher (Mentor)
Senior Product Designers x2 (SME)
Neural Divergent Group
UW MHCI+D Team
(Me) UX Designer + PM
1 UX Designer
1 UX Designer
1 UX Designer
My Impact
Designed four cognitive accessibility features with AI (Copilot) integration to reduce ADHD context switching, delivering concept sketches, user journeys, prototypes, and low/high-fidelity designs.
Conducted interviews with 3 subject matter experts and 5 neurodivergent users, synthesizing findings into 5 insights and 4 design principles that shaped the product strategy.
Led LLM system prompt refinement and evaluation, creating prompt guidelines and iterating testing to improve tone, structure, and usability.
Problem
For this project, it started with a brief to design an ADHD-friendly AI assistant, designed for professionals to support their daily workflow, built with Azure AI Foundry.
3.5%
🌎Across 10 countries met the criteria for adult ADHD
$105B~$194B
💸 Annual productivity loss per ADHD worker in the U.S.
22 days
📉 Lose more days of productivity/year than non-ADHD coworkers
With the right support, professionals with ADHD's unique superpowers can emerge. But unlocking those strengths also brings pain points to the surface.
Final Solution
Toolbox Widget
“Flow” is a toolbox widget that encomposes three key features (Magic Mouse, Activity History and Context Bookmarks) to help professionals stay in their workflow.
Design Recommendations
✔️ Be quietly available, never intrusive
✔️ Assist me like you know me
✔️ Let me practice my thoughts without pressure
✔️ Make engagement purposeful, not just playful
AI System Instruction (Neurodivergent-friendly)
I developed a neurodivergent-friendly AI system instruction using LLM GPT-4o model and Azure AI Foundry, iterating 4 times through structured user feedback. I directed the end-to-end evaluation cycle, laying the foundation for seamless integration with Copilot.
Developing custom system prompt on Microsoft's Azure AI Foundry environment
Key Features
01. Magic Mouse
01. Magic Mouse
Select & highlight any area on screen to instantly provide context to the AI, activating assistance with fast, context-aware suggestions without breaking your flow.
Area Select mode
Any area on screen can be selected to activate neurodivergent-friendly AI assistant. Each selection instantly becomes input, providing context-aware suggestions that deliver seamless assistance without disrupting workflow.
Text Highlight mode
Text-highlight is another way to activate the AI Assistant, designed for text-based tasks like writing or social scripting. By highlighting longer passages, users give the AI fuller context.
SOLVED PAIN-POINT
Fragile Momentum: With Magic Mouse, users can instantly activate AI assistance by turning any selection or highlight into immediate input, allowing them to stay in flow without breaking momentum or adding extra cognitive load.
76%
Surveyed participants (n=43) struggle with Feeling overwhelmed by task
“With ADHD, momentum is a big issue for me. When I open something and it’s just blank, it’s hard to even get started."
— P2, ADHD-diagnosed tech worker (Semi-structured interview)
SOLVED PAIN-POINT
➡️
I designed a custom AI Assistant tailored to the needs of neurodivergent users.
Explore the final AI system instruction process & its output here
Communication Block: The neurodivergent-friendly AI Assistant supports users by grounding them emotionally, reframing their perspective, and proactively anticipating needs before they are expressed.
56%
Surveyed participants (n=43) struggle with workplace communications
“I’ll get a little bit derailed by that and I'll overanalyze the words that were said. Gosh am I in trouble? Is this gonna reflect poorly on me? Am I about to get an angry message from my boss?“
— P4, ADHD-diagnosed tech worker (Semi-structured interview)
02. Activity History
Activity History passively tracks user activity through screen recording, creating a searchable timeline that helps users quickly revisit specific content.
Ask Copilot in Activity History to generate a smart summary of your past activity, revealing the why behind each step so you can instantly regain context and get back on track.
Copilot dynamically moves you to the exact point in the Activity History timeline based on your input, generates a smart summary for context, and lets you instantly relaunch the app file with a single click.
SOLVED PAIN-POINT
92%
Surveyed participants (n=43) answered staying focus requires significant effort
“Context switching is very painful... Getting back to a task after an interruption takes a lot of energy”
— P1, ADHD-diagnosed tech worker
(Semi-structured interview)
03. Context Bookmark
Users can capture content like highlighted text, screen snippets, links, and tools into focused Workspaces. From Workspaces, they can revisit saved items and set Smart Reminders on items to follow up on.
From Workspaces, they can revisit saved items and set Smart Reminders on items to follow up on.
SOLVED PAIN-POINT
Scattered Focus: Users can capture sparks of ideas in the moment and return later, reducing derailment from interruptions or new tasks. By keeping information visible, Context Bookmark anchors attention and preserves flow.
“Yeah, I I often find that hidden or information that's not immediately visible is hidden information for my brain. Like it's it just doesn't exist.”
— P1, ADHD-diagnosed tech worker
(Semi-structured interview)
“..I'll often like forget or get sidetracked with various asks for things that come up throughout the day.”
— P3, ADHD-diagnosed tech worker (Semi-structured interview)
Impact
🔹 Business Impact
Expected Business ROI Projection
+457%
Expected ROI with Flow integrated into Microsoft Copilot
Expected Time-Saving
+25 mins/day
Expected Time-Saving per user from productivity boost using Flow
🔹 User Impact
Adoption Usability
100%
Participants expressed interest in Adopting Flow into their workflow
AI System Prompt Satisfaction
4.3/5 Rating
Overall Satisfaction Score for custom AI Assistant (n=25)
🔹 Stakeholder Approved
“You should patent or publish paper on this. I’ve never seen anything like what you shared.”
— Microsoft Principal PM, Azure AI
⭐⭐⭐⭐⭐
“I think [the solution] resonates really well. [Regardless of ADHD,] we all need help for problems like this."
— Microsoft Senior Product Designer
⭐⭐⭐⭐⭐
RESEARCH
Learning about ADHD users
My Impact
Designed a survey and interview study guide to collect quantitative insights and recruit participants for in-depth interviews.
Led synthesis sessions on key trade-offs across product, design, and business goals, driving cross-functional alignment.
Conducted an internal seminar on Microsoft Azure AI to help the team design product-aligned solutions, improving engineering feasibility and stakeholder alignment.
Generative Research - What are the c
In partnership with Microsoft’s Neurodivergent ERG, we conducted three research methods to deeply understand the daily challenges neurodivergent tech workers face, what drives their motivation, and the tools and strategies they rely on. These insights directly shaped the design of a personalized AI agent to improve their focus, workflow, and overall well-being.
📍 See the detailed Research Report
📝 (n=43) Survey responses for the quantitative data and responses
🧠 (n=5) In-depth Semi-Structured Interviews to dive deeper into behaviors
🎤 (n=3) SME Interviews with Clinician, ADHD Coach, ADHD-productivity tool startup to understand existing solutions
The research led us to these key findings:
1
Social Communication: When professionals with ADHD face emotional overload or rejection sensitivity, they need a judgment-free space to pause, rehearse, and regain clarity.
2
Executive Function Management: ADHD professionals need systems that flex to their shifting focus and energy, because rigid tools often fail when their mental state shifts throughout the day.
3
Focus Management: Professionals with ADHD need support regulating the balance between hyperfocus and overwhelm before they lose momentum. Intentional rest, rather than increased effort, is key to sustaining productive flow.
4
Abrupt Challenges: Professionals with ADHD experience fragile momentum and benefit from support that minimizes disruptions, protects focus, and helps them quickly recover context so they can resume tasks without starting over or losing flow.
Internal Team Seminars: Azure Ecosystem & Azure AI Foundry
In parallel, I led Azure AI sessions to deepen the team’s understanding of the technical environment. These sessions aligned our cross-functional group on the system’s capabilities and constraints, enabling more informed, technically grounded design decisions.
📍I created a detailed Azure Cloud Basics Session & Azure AI Foundry deck to guide the team.
➡️Next step: Use these insights to ideate design solutions aligned with our defined direction.
IDEATION
Ideation by Key Research Insights
My Impact
Led ideation design workshop across product, design, and business goals.
Drove alignment on key design decisions across the team, streamlining collaboration and ensuring consistent direction that reduced misalignment and accelerated delivery
Based on insights from our research, I facilitated a design ideation workshop centered on the four key insights: social communication, executive function management, focus management and abrupt challenges.
Team working
🍭Tip: Always bring candy to your ideation - it makes you more creative🍬
➡️Next step: Conduct RITE design sprints to translate these insights into solutions.
DESIGN & ITERATIONS
6 Design Sprints with Rapid Iterative (RITE) Cycles
My Impact
Directed 6 RITE-driven design sprints, leading iterative cycles that accelerated problem discovery and solution refinement.
Explored 6+ feature concepts, evaluating tradeoffs and use-case relevance, then prioritized and scoped to the 3 core features.
Partnered with Microsoft Product Designers and recruited users to ensure the design direction was grounded in user needs and aligned with Microsoft’s design principles.
Overall process of the Design Sprints
Sketches to explore ideas
Mid-fi prototype (Static UI)
High-fi prototype (Interactive UI)
🔥 My role: I led the team in bringing structure and clarity to complexity.
I organized and led design workshops, guided team discussions, and consistently anchored design decisions in user research, technical feasibility, and the product vision to ensure we delivered impactful results within a tight timeline.
🚀 Impact: From scattered ideas to focused solutions.
Through synthesis and alignment, I transformed 6+ fragmented ideas into 3 validated core features. This gave the team clarity, focus, and a solid foundation for building the final product.
Flow Design System (based on Microsoft Fluent 2 Design System)
AI SYSTEM INSTRUCTION
Neurodivergent-friendly AI Assistant
My Impact
Directed the end-to-end evaluation of the system instruction, driving four user-informed iterations from concept to refinement.
Led the technical evaluation process, ensuring the system prompt aligned with Microsoft’s ecosystem standards, feasibility, and long-term integration.
Evolving AI System Instruction: 3 Iterations, 4 Versions
Overall process of the LLM System Instruction development
Iteration Comparison: AI System Instruction of Ver 1 -> Ver 4
Snapshots of System Instruction Evaluation Results & Process: Azure AI Foundry & Qualitative Evaluation & Synthesis
Multi-Agent Orchestration
Beyond Custom AI System Instruction: Scaling with Azure AI Foundry
My Impact
Turned technical orchestration into a product story (from single agent → scalable, specialized, production-ready system).
Built a demo to show enterprise value to pitch how one Custom AI Assistant could evolve into a modular platform aligned with product adoption within Microsoft Ecosystem.
Translated research into agent roles, creating user-friendly interactions for each agent that leads to enhance useability & inclusivity.
For Scalable, Enterprise Production-Ready Design: Split role into dedicated Agents
🤖 SINGLE AGENT
❌
Works for simple, isolated tasks: but falls short in real-world workflows that require multiple steps, decision-making, and context switching.
❌
Limited business value: lacks modularity, scalability, and flexibility for enterprise adoption.
🤖🤖 MULTI-AGENT ORCHESTRATION
✅
Scales horizontally: more agents can be added without disrupting the system.
✅
Enterprise-ready: modular design makes it auditable, governable, and adaptable to evolving workflows.
Demo: Multi-Agent Orchestration
Main Agent orchestrates Sub-Agents to keep work flowing
From the log, it shows Sub-Agent activates on demand to handle tasks dynamically - minimizing effort for the user
✅Scalability: Distribute tasks across multiple agents to scale horizontally with more users, workflow gets complicated.
✅Specialization: Each agent is fine-tuned for a specific role. This improves accuracy and efficiency, while making the system easier to build, test, and maintain within CI/CD pipelines.
✅Flexibility & Production Readiness: Agents can be reused and new agents can be added across workflows, supporting seamless deployment into enterprise environments.