Projects

I build AI systems that help people work with complex information, from natural-language tools for enterprise software and news archives to automated publishing workflows. This page also includes code released alongside my speech and audio AI research.

Industry projects

Hivedome Ltd.

2025 – present

ASK ITAS — AI Assistant for Commodity Trading

As a KTP Associate at Hivedome, I'm developing ASK ITAS, an AI assistant for ITAS, the company's commodity trading and risk management platform. The first release supports natural-language questions over product knowledge, with answers grounded in referenced sources.

I'm extending it towards an agentic workflow that can query live Data Portal records and ask clarifying questions when additional context is needed. These tool-driven capabilities are in development.

LLM Agents · RAG · Tool Use · Enterprise AI

Eidosmedia

2023 – 2025

Question Answering Over Newspaper Archives

Built a retrieval-augmented question-answering assistant for editorial teams using large newspaper archives. The system retrieves relevant articles and returns answers with source citations. I contributed retrieval design, news-domain language-model adaptation, and AWS-based training and deployment pipelines.

RAG · Semantic Retrieval · NLP · AWS

Eidosmedia

2022 – 2024

Autonomous Newspaper Pagination

Developed a reinforcement-learning system for arranging newspaper articles into page layouts based on an editorial plan. The project began as my MSc thesis at the University of Bologna and was subsequently developed into a production system at Eidosmedia.

The work included assembling a corpus of published newspapers for training and evaluation, and developing the models used to make layout decisions under the constraints of the editorial plan.

Reinforcement Learning · Dataset Development · Production ML

Research implementations

  • CARD Training, distillation and inference code for encoder-free audio captioning, along with model resources and checkpoints. PaperDemo
  • SpeechLLM-Pruning-Fairness Experiments for speech-encoder pruning and demographic-group evaluation using Fair-Speech and Common Voice. Paper
  • SLAM-ASR-Encoder-Pruning-LoRA Code and configurations for Whisper encoder-depth experiments, projector training, and LoRA adaptation in SLAM-ASR. Paper