Millions of scientific articles are published every year. In the arXiv repository alone – which is dedicated to computer science and related disciplines – there are over three million entries. For any researcher, finding relevant information amongst such a vast amount of data is a challenge that goes beyond human capacity. Trenton Fowler, recognised in the Information and Communication Technologies category at the EDUCA EDTECH 2026 Foundation’s Academic Excellence Awards, decided to develop a solution.
The problem he set out to solve
Information overload in academic research is not a new problem, but it has worsened exponentially over the last decade. Traditional search engines return results based on keywords, not on meaning. A researcher seeking answers to a complex question needs something more sophisticated: a system that understands what they are asking and retrieves what really matters.
That is exactly what Trenton built: a document intelligence platform capable of analysing 50,000 scientific papers from arXiv and answering questions in natural language with cited and verifiable answers.
How it works
The system’s architecture combines several cutting-edge technologies into a coherent workflow. The process begins with Apache PySpark, which filters, deduplicates and cleans more than three million records to produce a sample of 50,000 articles across ten subcategories of computer science.
The abstracts of these articles are converted into 384-dimensional mathematical vectors using a Sentence Transformers model, generating more than 164,000 fragments indexed in a ChromaDB vector database. In parallel, the structured metadata is stored in PostgreSQL for relational queries and statistical analysis.
The web interface, built using React, offers two main features. The first is an analytics dashboard that displays publication trends, distribution by category, author rankings and a two-dimensional visualisation of the topics in the corpus. The second is a Retrieval-Augmented Generation (RAG) system which accepts questions in natural language, retrieves the most relevant fragments based on semantic similarity, and generates cited responses using the Mistral 7B model running locally.
The results
The figures speak for the system. In an evaluation based on 30 curated questions, the retrieval component achieved an accuracy of 75.3% and a recall of 92.8%. The generated responses achieved an average fidelity of 4.57 out of 5 and a relevance score of 4.63 out of 5. The retrieval latency averaged just 38.5 milliseconds.
One particularly important detail: the entire system runs locally, without relying on cloud services, packaged into five Docker containers that are started with a single command.
Why it matters
Trenton’s work is not merely a technical exercise. It is a practical demonstration that artificial intelligence tools can be put to use in scientific research in a way that is accessible, reproducible and transparent. At a time when the debate on AI in academia is more heated than ever, his platform offers a concrete answer: technology can help researchers find what matters, without replacing their judgement.
