Work

AI-powered pipeline for generating European project proposals

Year: 2026
Type: dissertation
Abstract: First of all, I would like to mention that this thesis has been developed in collaboration with Sparsity, a technology company and spin-off from the DAMA-UPC (Universitat Politècnica de Catalunya) research group, with expertise in data management and advanced algorithmic design for mobility. I would especially like to thank my thesis director, Josep Larriba Pey, for giving me the opportunity to carry out this thesis within his organization and for the time he dedicated to discussing the direction of the work with me. The initial idea and motivation behind this thesis work was to be able to generate European project proposals through an AI-powered pipeline. But I soon realized that trying to train an LLM model with such a limited corpus was non sense. The same problems arise when trying to apply some kind of fine-tuning. And doing just prompt engineering with a pretrained LLM model was not interesting enough from a scientific research perspective. So, it was in this moment that I realized that the real value of the corpus was in the ESRs (Evaluation Summary Reports). Horizon Europe is the largest transnational competitive research funding program, yet acceptance rates in many calls is quite limited. Rejected consortia receive Evaluation Summary Reports with detailed textual feedback, but no systematic method exists to extract, structure, and analyze this content. The feedback stays locked in PDFs, and they are usually read once by the company supervisor. So, I decided to develop, apply and validate a four-phase analytical framework that turns unstructured evaluation feedback into structured data, building a Knowledge Base, which it can be used as a support system in writing a new proposal. The pipeline has four phases: LLM-assisted extraction of evaluative sentences, semantic gap analysis linking criticisms to proposal text, alignment testing between proposals and call documents, and retrieval-augmented recommendation generation. Applied to 18 Horizon Europe projects (3 accepted, 15 rejected) from a Sparsity’s portfolio, it produces 896 evaluative sentences classified by polarity and category. The framework is then deployed as Calliope, a web application that turns the methodology into a production-ready service. After running the pipeline, three important findings emerged. First, more than 90% of the 309 negative criticisms target content that is present but inadequately developed. This contradicts the expectation that rejected proposals fail due to missing topics. Accepted proposals also write more on the same criticized topics. And finally, textual alignment between proposals and call documents does not predict acceptance. The null result shaped the recommendation engine, which treats alignment as a coverage diagnostic rather than a quality predictor. After the analysis was performed, a web application called Calliope was developed to apply the extracted knowledge and to provide a stable, production-ready service for the Sparsity organization. In this way, the organization can benefit from the resulting knowledge base, using it as a proposal repository, a proposal-writing support system, or even a proposal draft analysis tool. It is also worth mentioning that an ingestion feature was implemented to allow the incorporation of future proposals into the knowledge base, an essential capability. Over time, this will enable the organization to benefit from new Horizon project attempts by expanding the corpus and making the analysis more robust.
Language: English
Open Access status: green
OpenAlex ID: W7163778626
Cites: 30
Cited by:
Related to: 10