CiteMed · Elisey Kolev

Restructured an AI-assisted clinical-evidence platform around how medical device teams review literature, and cut screening time by 50%.

Role: Senior Product Designer
Timeline: Feb – Jul 2026
Shipped
Scope: Product strategy
Information architecture
UX/UI
Worked with: CEO
Product
Engineering
Industry experts
Industry: Healthcare technology
B2B SaaS

Overview

One connected experience for the entire literature review

CiteMed helps medical device teams find, evaluate, and organize evidence from large amounts of scientific literature, and turn it into structured clinical evidence for regulatory submissions.

Systematic reviews mean working through thousands of papers to identify relevant studies, extract key information, and compile a final report. The process is rarely linear: teams revisit searches, update criteria, and review earlier decisions as their research develops. Their team brought decades of industry experience, but as the product grew feature by feature, connecting these steps became harder. CiteMed's aim is to bring the whole process together in one place, from the first search to the final report.

CiteMed reached out to improve the information architecture and make the product more workflow-driven. Working closely with the CEO and industry experts, I restructured it around the stages of a literature review and redesigned search, screening, evidence extraction, and the reference library, with contextual AI built into each.

10+ → 4
More than 10 navigation options consolidated into four stages of a literature review.
50%
Less time spent screening scientific literature.
AI built in
Assistance in search, screening and extraction, reducing manual effort.
Reusable research
A shared reference library teams build on instead of starting from scratch.

“Elisey has done terrific UX and UI work for us. We operate in a complex industry, with incredibly technical users. He's made an involved workflow substantially more streamlined for our users, and has led our team in re-imagining what's possible for our user experience. Couldn't recommend him enough.”

Ethan Drower Founder & CEO, CiteMed

Design principles

Follow how reviews really happen
Teams move back and forth between stages, so the product never forces a fixed sequence.
AI suggests, researchers decide
Recommendations, extracted data and multi-step actions all wait for the researcher's approval.
Every value traces back to its source
Extracted evidence stays linked to the passage in the article it came from.

/01. Information architecture

From scattered features to a clear workflow

I started with detailed conversations with Ethan Drower, CiteMed's CEO, to understand how medical device teams approach literature reviews and where the product structure created friction.

Protocols, searches, screening and evidence review were spread across more than 10 navigation options. Each supported an important step, but users struggled to see where they were and what to do next. The navigation also treated reviews as linear, when in reality teams keep returning to earlier stages to refine criteria, update searches and review more studies.

Before and after: the old product with a long sidebar of sections such as Search Protocol, Search Terms, Run Searches, 1st Pass Review, Adverse Events, 2nd Pass Review and Finalize, followed by the redesigned screens — Before & after.

I restructured the product around four core stages, with a home view for what needs attention. Each stage answers one question researchers ask, and users can move between stages as their research evolves.

From scattered features to a clear workflow. Home: what needs my attention? See recent activity, pending work, and quickly get back into the review. Protocol: what are we trying to find, and how are we defining it? Manage the research approach, screening rules, extraction fields, and other review methodology in one place. Screening: what should we search for, and which studies are relevant? Define and run searches, then review the results in the same workspace. Full-Text Review: what does each study tell us? Review full articles, extract structured data, and compare studies. Reports: what can we conclude from the evidence? Turn the research into reports and meaningful outputs. — After: more than 10 navigation options became four stages and a home view.

From there, I designed the end-to-end flows across every stage: configuring protocols, defining searches, screening articles, full-text review, managing columns, and the reference library with its summarize, extract, share and smart list flows.

Figma overview of the redesign's flows, grouped as Configure Protocol, Define Search, Screen Articles, Full-Text Review, Manage Columns and Reference Library, with Reference Library flows for Summarize, Extract, Share List, Natural Language Prompt, Generate Smart List and Smart List — The scale of the redesign: the flows in Figma, grouped by stage.

/02. Search and screening

Better searches, faster decisions

The quality of a review depends on its search. Too narrow, and it misses important evidence. Too broad, and it returns thousands of irrelevant articles that still need screening.

Starting from a research question

Database queries use specialized syntax that takes real experience to write. Feedback from early users showed that many struggled to turn research questions into queries that produced useful results. I designed a guided search that starts with a question in plain language, suggests what would make it stronger, such as specificity, outcomes of interest or comparative context, then generates database-ready syntax.

Start with a question dialog over the list of searches, rating the research question and suggesting specificity, outcomes of interest and comparative context — Starting a search query with an optimized question.

Full control for experienced researchers

Experts still want to write their own strategy. I organized the query builder around PICO, a widely used framework for clinical research questions, with search terms suggested from the protocol and device details. It validates syntax in real time, so errors are caught before a search runs.

Query builder with PICO blocks, suggested search terms, and syntax errors and warnings with a Fix Error action — Query builder with syntax validation and suggested search terms.

Screening built for quick decisions

After a search, researchers decide which of hundreds of articles move forward. I kept abstracts visible next to the screening actions, so studies can be judged without opening and closing records. AI Screening Assist highlights relevant passages and recommends a decision based on the protocol criteria.

50% less time spent screening literature.

Screening table with Retain, Maybe and Exclude actions beside each abstract, and AI Screening Assist recommendations with scores — Screening article abstracts, with AI Screening Assist recommending against the protocol criteria.

/03. Full-text review

Reviewing studies without losing the whole picture

The previous experience showed basic article information in a table, with little sense of progress. Opening an article led to a long list of extraction fields, making it hard to understand the study as a whole.

Finding specific details in a paper, understanding them in the context of the study and recording them consistently takes significant time. I wanted to make that work easier without taking researchers out of the review.

I organized extracted data into study arms and structured sections, reflecting how researchers break down evidence. AI-assisted extraction finds the relevant information in the article and gives researchers a starting point to review.

Extraction form organized by study arm, with Show in paper links beside the article and Reject or Save and Approve controls — Extracted data is organized by study arm, and every field links back to the paper.

Comparing evidence across studies

Across the full set of studies, researchers shape their own view of the evidence, choosing the fields they need to compare instead of working from fixed columns.

Full-text review table with chosen columns such as study design and sample size, and a side panel highlighting the source passage — Researchers pick the columns that matter for their analysis, such as study design and sample size.

/04. Smart reference library

Research that teams reuse across devices

The same literature often supports several device reviews over time, but references were tied to individual projects. I moved them into a shared library that teams build on and shape around what they need to compare.

AI makes the library useful beyond what's already captured: researchers can ask new questions across their references and see the answers as columns.

Reference library with a menu for creating an AI-powered column from a plain-language description — Creating a custom AI-powered column.

Smart lists extend this further. Teams describe what they're looking for in plain language, refine it with AI and structured rules, and the platform keeps the list up to date as new research is added.

Smart list builder asking what the list should include, with a plain-language description, AI rules and structured rules such as population, setting, publication type and study type — Configuring a smart list in the reference library.

/05. Clio AI assistant

AI that works alongside researchers

A generic chatbot wouldn't fit a complex evidence workflow. Researchers need help that understands what they're working on.

I designed Clio, a contextual assistant that adapts to where researchers are: an article, the reference library, or a report. It summarizes information, extracts data, finds studies, organizes references and improves content.

Report editor with the Clio Assistant panel proposing edits that the researcher can accept or discard — Clio suggesting edits in a report, which the researcher accepts or discards.

Clio can also complete multi-step actions, such as finding studies, creating a list and sharing it with the team. It keeps researchers involved by showing the planned steps and asking for approval before making changes.

Clio Assistant in the reference library showing a three-step plan: Find Studies, Create List, Share List, with the list title waiting for approval — Clio performing a multi-step action, one approved step at a time.

Outcomes

From a collection of features to one workflow

The redesign brought CiteMed's capabilities together around the way systematic reviews actually happen, helping researchers understand where they are, make better decisions, and move through complex evidence workflows with less friction.