I am Maurizio Scibilia, an AI Engineer focused on retrieval, evidence grounding and AI evaluation. I design systems where failure modes can be analysed, benchmarked and improved rather than hidden behind an end-to-end result.
Local-first evidence retrieval. ParancU explores an evidence-first approach focused on finding answer-bearing passages rather than merely semantically similar text.
The project grew from the EcoSearch v1.0 research prototype and was carried across the full AI system development cycle: research framing, architecture, implementation, benchmark design, failure analysis, mobile integration and App Store distribution.
The current implementation keeps corpus persistence, multilingual E5 inference, scoring, ranking and evidence inspection on-device. Retrieval is question-oriented, with bounded lexical evidence available when exact or rare terms matter.
The goal is not a black-box demo. ParancU is designed as an inspectable body of work, with documented architecture, evaluation results and a white paper that makes both the contribution and the remaining limitations explicit.
The project evolved by questioning the retrieval objective itself.
EcoSearch began as a retrieval-first QA prototype designed 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: semantic proximity is useful, but does it reliably identify the passage containing the evidence needed to answer a question?
ParancU focuses directly on that retrieval problem, independently from downstream language generation.
Experience across software engineering, consulting, web development, data science and AI evaluation, including work with Siemens, Framfab / A.T. Kearney Applications & Technology, Deep Consulting, Spindox and others.
Hands-on experience designing and building technical systems across different generations of software, from web and application development to modern AI and retrieval pipelines.
Current work places particular emphasis on benchmark design, failure analysis, retrieval quality and the distinction between plausible output and inspectable evidence.
University of Pisa · A (highest honours). 18-month research thesis on LRAM graph neural networks, including original theorems on asymptotic and exponential stability.
Master in ICT, Cefriel / Politecnico di Milano · A · Scholarship recipient.
MBA, LUISS Guido Carli · A · Scholarship recipient, with international exchange at Paris Dauphine University.
2021–2024
2023–2025
AI engineering with a strong emphasis on retrieval quality, evidence visibility and measurable system behaviour.
My work combines research framing with hands-on engineering: defining the problem, building the system, constructing evaluations, analysing failure modes and iterating until the behaviour becomes measurable and useful.
For full professional background, education, certifications and contact details, see my CV.