ROLE
Design Engineer • Product Designer
Open to new roles
2026
CASE STUDY 02 — AGENTIC EXPERIENCE
AI GRANT COPILOT
An AI copilot that guides researchers from raw idea to a compliant submission, where the agent does the labor and the researcher stays the accountable decision-maker.
TIME
2025
CRAFT
Agentic UX • Design System • Coded Prototyping • Trust & Safety
TOOL
Figma • Claude
grant-copilot / match
View Prototype↗
THE CHALLENGE
Funding research shouldn’t mean fighting the paperwork.
Finding the right opportunity means navigating scattered systems, decoding dense eligibility rules, and racing deadlines - mostly by hand.
01
Scattered across systems
Opportunities live across separate federal portals and mailing lists - none sharing state. Researchers stitch together sources by hand just to see what’s open.
02
Relevance is manual
Thousands of solicitations, no real fit ranking. Finding the few that match your methodology means reading dense PDFs one by one.
03
Eligibility hides on page 40
Up to 30–40% of applications are disqualified on technical eligibility — criteria buried deep in the solicitation, surfaced too late to act on.
The System
A four-step pipeline that makes a complex application legible.
IN
Pitch in
one raw research pitch · PDF or DOCX
HUMAN GATE
04OUT
Package out
one submission-ready application ✓ · human-approved, exported
Intake
Upload your pitch — the agent plans before it executes.
pitch.pdf→agent plan
I replaced a 15-field search form with a single document upload. The agent reads the applicant’s intent, then outlines the steps it plans to take before it executes, so automation begins with a shared plan, not a black box.
Intent-drivenPlan before execute
STEP 01 / 04 · INTAKE — Upload → the agent proposes its plan and asks to proceed.
Match
Review ranked matches — and see why each one fits.
agent plan→ranked matches
Dense model analysis and cross-referenced institutional data become clear, ranked, actionable matches - communicating why each one fits, not just that it does
Explainable matchesFit scoring
STEP 02 / 04 · MATCH — Ranked matches, each with a fit score and the reason behind it.
Eligibility
Verify eligibility yourself — the agent only flags.
ranked matches→verified match
Compliance documents and applicant status are verified with a human review gate at every high-stakes step.
Human review gateCompliance check
STEP 03 / 04 · ELIGIBILITY — The agent flags what’s missing — you stay the one who verifies.
Package
Assemble, approve, export — as one package.
verified match→submission package
Every section is assembled, reviewed, and exported as one submission-ready package.
One-click assemblyExport-ready
STEP 04 / 04 · PACKAGE — A live readiness meter — each section approved before the package can export.
The Signature Design Decision
Review before you Confirm
The first build had a single “Confirm” button that asked applicants to certify a compliance document they’d never seen. One critique named the problem precisely: “Confirm is a terminal verb - it implies you’ve already reviewed.” The fix wasn’t a label change. It was a structural rethink.
Before — v1 · One-click confirm
Current & Pending Support · Action Needed
Confirm
The user is asked to certify a disclosure without seeing its contents. One click marks it “Ready.” Nothing to inspect. Nothing to dispute. The mental model: trust us, it’s fine.
What goes wrong
An applicant certifies that an award is still “Pending” — but it was funded two months ago. The reviewer flags the misrepresentation. The submission is voided.
Certify without seeing
No disagree path
Trust smell
After — v2 · Review → Confirm loop
grant-copilot / review
The compiled disclosure is visible before any decision. Confirm & attach is irreversible — it lives inside the drawer so the verb is honest. Request changes gives disagreement a first-class, low-friction path.
Review opens content
Confirm lives inside
First-class disagree
Revise & re-review
HOW IT WAS BUILT
Designed and shipped with an AI co-pilot.
This wasn’t just a design about AI — it was designed with AI. I ran an AI-assisted design-engineering workflow end to end, from system architecture to a coded, interactive prototype.
01
Figma
Strategic Blueprinting
Architected an AI-native design system and accessible component tokens - structured deliberately so each component was directly consumable by an AI coding agent.
→
02
CLAUDe
Prompt-Driven Prototyping
Fed structural parameters and design intent into an AI co-pilot to rapidly translate static design into a functional HTML/React front end - real components, real state.
→
03
the roi
10× Faster Iteration
Tested real interactive states and edge cases directly in the browser, collapsing the concept-to-delivery loop from weeks to days.
OUTCOMES & TAKEAWAYS
On high-accountability products, the best design decision was adding a step, not removing one.
impact
A fragmented, multi-portal ordeal became a single guided pipeline where an AI agent does the labor and the user keeps the decisions with zero blind certifications by design.
Process
I mapped the real workflow before touching UI, iterated through a one-click first draft, caught its trust gap in critique, and rebuilt around a Review → Confirm loop - later systematized into one reusable component.
What I learned
Owning both the design system and the coded prototype let me design for an AI agent and partner with one to build - proof that hard constraints don’t limit good design, they tell it exactly where to put the human.