ScienceScan.ai

Inspiration

Personally I always found it difficult to decipher the countless troves of scientific publications. Springer Nature, The Cell, Lancet etc. all host world class publications but there is no singular way to search based in a streamlined fashion based on a topic across all journals. While a couple exist, quality is lacking and LLM hallucination is real. I also wanted a tool for personal use to stay relevant with recent publications and this provides that.

What it does

ScienceScan strictly pulls from verified sources, i.e. published scientific literature from top journals. User-friendly, query-based natural language input for the right literature. Providing consumers with easy access to verifiable literature in a digestible format and avoiding LLM hallucination. Using RAG methodology from databases such as Crossref, we are able to streamline the "staying up to date" for professionals, researchers, and clinicians.

How we built it

Utilizing Crossref API for relevant publications procurement; OpenAI API for LLM generation; Streamlit for python-based front-end and prototype deployment.

Challenges we ran into

Doing this project solo with little full-stack development experience proved difficult at times. The biggest issue was navigating the Crossref database effectively. While there are over 130+ million records, missingness is still a problem as well as varying identifiable keys: the Spring Nature journal is recorded as 'Springer Science and Business Media LLC' but also in other languages and with different jargon.

Accomplishments that we're proud of

As a first time hackathon attendee, I am proud of spearheading this project by myself, by learning full-stack development and deployment on my own and creating something that I personally have a use case for.