Strawberry Matcha
AI agent for marriage-based green card applicants (CR1 or F2A) filing without a lawyer.
- Role
- AI UX Designer · Solo project
- Tools
- Cursor, Claude API, Supabase, Figma
- Focus
- Conversational AI, Decision-support UX
- Timeline
- 3 weeks · v0 → v1
Problem
Filing alone leads to mistakes. General AI makes it worse.
Many couples applying for a marriage-based green card file without a lawyer. Legal fees run thousands of dollars, and the process looks doable, so they handle it themselves. Then the details catch up. 1 in 4 applicants gets a Request for Evidence for avoidable errors, and each one adds three to five months. General AI doesn't fill the gap. It hallucinates on legal details and answers for a generic case, not theirs.
Key features
Ask Strawberry Matcha, a conversation that knows your case.
Users can ask anything, anytime. Strawberry Matcha answers based on the applicant's actual case status and preparation progress, and updates the case as the conversation continues.
Field Translator, fills the gap between your real life and the form.
When users upload any edition of a USCIS form PDF, Strawberry Matcha reads the actual form fields, cross-references them with the user's case data, and tells them exactly what to enter in each field. It also handles tricky format conversions, such as restructuring a Korean address to fit U.S. form fields or matching a Korean name to its passport romanization.
User input
Free-form Korean address as the applicant naturally writes it.
USCIS form output
- ProvinceSeoul
- City or TownGangnam-gu
- Street NameTeheran-ro
- Street Number123
- Apt / UnitDong 101, Ho 202
Parsed and reformatted into the exact fields each USCIS form expects.
Timeline guidance, so you know where you are and what's next.
Each milestone shows where the applicant is in the process, what the step actually means, and what usually happens next, so the case never feels like a black box.
How I Built
From concept to crafted product in five steps.
- 01Domain research
Define concept & Research to train the AI
Mapped how immigration lawyers actually walk a couple through CR1 / F2A.
- 02Cursor plan mode
Design System Architecture
Used Cursor's plan mode to map out the full system as a diagram, so I could see how every piece fit before writing code.
- 03Cursor prototype
Fast validation
Used Cursor to spin up a working prototype quickly, so I could test the idea with real applicants before investing more.
- 04Real applicants
Iterations
Reworked chat structure and onboarding based on where trust was breaking.
- 05Figma polish
Craft refinement
Polished the UI in Figma, tightening tone, pacing, and visual hierarchy across the whole product.
- 01
Domain research
Define concept & Research to train the AI
Mapped how immigration lawyers actually walk a couple through CR1 / F2A.
- 02
Cursor plan mode
Design System Architecture
Used Cursor's plan mode to map out the full system as a diagram, so I could see how every piece fit before writing code.
- 03
Cursor prototype
Fast validation
Used Cursor to spin up a working prototype quickly, so I could test the idea with real applicants before investing more.
- 04
Real applicants
Iterations
Reworked chat structure and onboarding based on where trust was breaking.
- 05
Figma polish
Craft refinement
Polished the UI in Figma, tightening tone, pacing, and visual hierarchy across the whole product.
Iterations
Reducing cognitive overload in chat.
The first version dumped each response into one long paragraph. In user testing, the answers were accurate but people didn't act on them. They skimmed, asked me to repeat things the AI had already said, and gave up mid-task. I explored three response formats before settling on one.
Version 1Single response paragraph.
- Fast to implement; no extra UI.
- Buries what matters most.
- Users don't know what to ask next.
Full doc-style hierarchy with headers and bullets.
- Maximum scannability.
- Loses conversational warmth.
- Overkill for short answers.
Two-layer voice (serif acknowledgment + sans-serif info) with suggested follow-ups.
- Reads warm and human.
- Scannable at a glance.
- Nudges the next question.
- More design and prompt work.
Single response paragraph.
- Fast to implement; no extra UI.
- Buries what matters most.
- Users don't know what to ask next.
Full doc-style hierarchy with headers and bullets.
- Maximum scannability.
- Loses conversational warmth.
- Overkill for short answers.
Two-layer voice (serif acknowledgment + sans-serif info) with suggested follow-ups.
- Reads warm and human.
- Scannable at a glance.
- Nudges the next question.
- More design and prompt work.
Final pick:
Version 3
Two-layer voice keeps the chat warm but makes the answer scannable, and the suggested follow-ups stop users from getting stuck on what to ask next.
Redesigning onboarding to stop hallucinations.
The original onboarding was too short. The AI guessed to fill gaps, and hallucinations broke trust fast. So I studied how immigration lawyers intake clients. Their upfront questions are how a lawyer learns the case before giving advice. I rebuilt onboarding around those same questions, so the AI starts with enough context to be accurate from the first message.
Reflection
What I took away from designing an AI agent for a high-stakes legal workflow.
Designing an AI agent is designing how it thinks.
Most of the work happened underneath the screens. Prompts, follow-up logic, what the AI asks versus what it answers, what it stores about the user. The visible UI was the smallest part.
Onboarding is data acquisition, not a signup.
How well an AI agent performs depends on what it knows going in. Designing onboarding well is designing the AI's first impression of the user, and everything downstream flows from there.
Conversational UX is about pacing, not just tone.
Users filing alone don't need more information. They need information at the right moment, in a shape they can act on, with a clear next step. That's a design problem, not a content problem.