DataExtractionSystemForLivingSystematicReviews

LLMs can help automate extraction. But there was no interface that made it work for researchers. We built one.

Client

Mayo Clinic

Timeline

4 months

Industry

Healthcare Research

Type of Work

Product Design | AI Interface

Backgroud

Manual.Time-consuming. Prone to error.

Researchers at Mayo Clinic were copying values from PDFs into spreadsheets by hand. Across hundreds of studies. Every time a new paper dropped. We set out to fix that.

The Problem

3 big problems.

Manual

Researchers copy values from PDFs into spreadsheets by hand. Every time a new study drops, the cycle starts again.

Time-consuming

A single LSR can take months. Existing tools add configuration overhead instead of reducing it.

Prone to Human Error

Manual transcription introduces missed fields and undetected conflicts across large teams.

Why not just use ChatGPT? The tools exist, but the interface doesn't. That's the gap we designed for.

The Process

how we got there.

"Can we make it simpler without losing control?" If not, we cut it.

Brainstorming

Mapped the full LSR workflow and surfaced the highest-friction moments researchers actually face.

Affinity Diagramming

Clustered pain points to find the design priorities that mattered most before touching a wireframe.

Lo-Fi Wireframes

Rapid sketching of layout flows and interaction models. No commitment to visuals, just structure.

Hi-Fi Prototype

Fully realized interface with real data structure, component library, and every decision justified.

Design Process

Task Flows

Version1Version 1 - initial task flow mapping core paths from dashboard through review creation, extraction setup, AI model selection, and table configuration.

Dashboard

Entry point

→

Create Review

Name + invite members

→

Settings Review

Configure extraction

→

Start Extraction

AI begins parsing

→

Tables + Figures

New form or template

→

Edit + Review

Human verification

→

Pass for Review

AI Build or Human Build

Version2Version 2 - refined flow with conflict detection, AI confidence branching, human override paths, side accept/reject, and a streamlined export step.

Dashboard

Entry point

→

Setting-up Extraction

Name, invite, link, AI model

→

Files Uploaded

Review the set-up

→

Start Extraction

AI parsing begins

→

AI Parsing

Confidence scoring

→

Extraction Review

Alter tables or figures

→

Review + Export

CSV, JSON, Excel, API

Version 2 cuts the ambiguity from Version 1. Every branch has an exit. Every AI step has a human fallback.

The Product

Manual.Time-consuming. Prone to error.

01 — Dashboard

all your reviews. one scroll.

Active reviews, conflict counts, and recent work front and center. Upload to start an extraction immediately.

Quick access

Customize Dash

Group your reviews into separate projects

🖥

Add screen image

The Product

Manual.Time-consuming. Prone to error.

01 — Dashboard

all your reviews. one scroll.

Active reviews, conflict counts, and recent work front and center. Upload to start an extraction immediately.

Quick access

Customize Dash

Group your reviews into separate projects

The Final Product

The FinalProduct.

How did designing for AI change the way I think about UX?

Working on this project pushed me to rethink the role of a designer in systems where AI is not just a tool but a collaborator. I came in assuming AI would simply automate the repetitive work. What I found was more nuanced.

Embedding LLM capabilities into a workflow only works when users can understand, trust, and know when to override what the AI produces. That meant designing for transparency at every step. Confidence scores, conflict flags, source traceability, these were the design, not features bolted on at the end.

This project made me a more careful systems thinker, a stronger advocate for progressive disclosure, and much more intentional about where human judgment should never be replaced.

Hero

Background

Problem

Before

Process

Metrics

Screens

Brand Colors

Logomark

Typography

let's talk!

Hop on a Call

Copy Email

back to portfolio

←pri.anka

next project

WalletGyde→

Product Design