2026
AI assistant for
oncologists
oncologists
about
This service was built during the Digital Health hackathon, organized by the Petrov Oncology Center and ITMO University. Its purpose is to help oncologists find relevant scientific materials. As a designer, I first designed the full functional system, then identified the MVP I could deliver within the deadline – and built a working solution based on AI agents.
Project specifics
Oncology is one of the most research‑intensive fields – and the volume of scientific publications is enormous: studies, reports, guidelines, and more. In Russia, research is ongoing, but to stay fully current, practitioners need access to international databases. Most of these – including major aggregators like Semantic Scholar – are currently unavailable in Russia.
Medicine in general and oncology in particular are highly sensitive to information quality – decisions are made based on this data, so any kind of hallucinations or made‑up content is unacceptable.
Even before the design work began, I had to define a few key constraints that shaped the entire architecture of the project.
Medicine in general and oncology in particular are highly sensitive to information quality – decisions are made based on this data, so any kind of hallucinations or made‑up content is unacceptable.
Even before the design work began, I had to define a few key constraints that shaped the entire architecture of the project.
Database availability
I settled on PubMed – an accessible database of over 40 million biomedical and biological articles, maintained by the U.S. National Center for Biotechnology Information (NCBI). It also has a well‑documented API, which makes integration straightforward
Query interpretation
Many scientific databases work with optimized search queries – and PubMed is no exception. It even has a standardized thesaurus: MeSH. But clinicians who don’t use research databases regularly may not be familiar with this – and may not have sufficient English proficiency either
Verifiability
Doctors need to trust both the search results and the summaries the system provides. All references must be valid and, where possible, link to freely accessible full‑text materials. The user should always understand the reasoning behind the system’s output – and the value of the sources it returns
Process
- 01Studied the task and defined key constraints
- 02Developed the overall technical concept and architecture
- 03Developed the UX concept and created the initial design
- 04Identified the MVP and defined
the tech stack - 05Built and tested AI agents, wrote code, and handled frontend
- 06Created the presentation and recorded the project defense video
Building a search system that summarizes scientific articles — while keeping hallucinations to a minimum — is a real challenge. I solved it by splitting the process into discrete, predictable stages, with different AI agents handling different tasks. This not only improved accuracy but also kept the cost per query down to just 20 RUB — even with YandexGPT Pro used across all agents.
Core logic
The MVP consists of the following modules.
Tech stack: Docker | FastAPI | YandexGPT Pro (API) | Python | HTML/CSS + JS
For my own research purposes, I’ve also built a version using OpenAlex — a broader database not limited to biomedical topics. I’m also planning to migrate to Hugging Face. The core logic and functionality of the service will remain unchanged.
Tech stack: Docker | FastAPI | YandexGPT Pro (API) | Python | HTML/CSS + JS
For my own research purposes, I’ve also built a version using OpenAlex — a broader database not limited to biomedical topics. I’m also planning to migrate to Hugging Face. The core logic and functionality of the service will remain unchanged.
Outcome
From a UX perspective, I wanted to be transparent with the user — showing exactly how many sources the summary is based on. Since at the MVP stage we only analyze abstracts for speed, the system clearly indicates that as well
Design system
The design concept leans heavily on Material Design 3, as well as established UI practices common in popular AI interfaces. But there are a few nuances.
