I am Maurizio Scibilia, a Data Scientist working across information retrieval, NLP, local AI, evaluation, and question answering. My current work centres on ParancU, an evidence-retrieval system that explores alternatives to conventional query-to-chunk similarity.
Dense retrieval is useful, but semantic proximity is not the same thing as evidence containment. ParancU starts from that distinction.
Instead of relying only on similarity between a user question and document chunks, ParancU compares questions with question-oriented retrieval metadata associated with source passages. Bounded lexical evidence can also influence ranking when exact or rare terms matter.
The current system is local-first: corpus persistence, multilingual E5 embedding inference, scoring, ranking, and evidence inspection run on-device. Question-oriented retrieval metadata is generated during corpus preparation.
ParancU grew from a broader investigation into retrieval quality, evidence visibility, and grounded question answering.
EcoSearch began as a retrieval-first QA prototype built to make supporting passages visible before generation. It became a test bed for chunking, semantic search, lexical signals, reranking, confidence routing, and retrieval evaluation.
Those experiments exposed a more fundamental question: should document retrieval depend primarily on similarity between the wording of a query and the wording of a passage?
ParancU narrows the problem. Its focus is not general chatbot generation, but the retrieval layer itself: whether the system can surface the specific evidence required by a question.
This makes retrieval behaviour easier to inspect, benchmark, and reason about independently from downstream language generation.
I am particularly interested in systems where model behaviour can be measured, inspected, and improved rather than treated as an opaque end-to-end result.
My work combines data science with hands-on system design: defining a problem, building the prototype, constructing evaluations, analysing failure modes, and iterating until the behaviour becomes measurable and useful.
For professional background, experience, and contact details, see my CV or GitHub profile.