Daniele Comi

Manager, AI Engineer

Technical lead on generative and agentic AI pursuits: discovery with the client, solution architecture, executive pitch, then staying hands-on through delivery.

• Architecting agentic and retrieval systems for defence and justice-sector clients, including fully air-gapped deployments on OpenShift with watsonx, Granite and Milvus.
• Built an LLM red-teaming practice on the ARES framework, with an OWASP LLM03 supply-chain attack plugin and a CycloneDX SBOM pipeline for AI components.
• Reference AI architectures for IBM Consulting Advantage, proposed to enterprise clients across banking, energy and the public sector.
• IBM Redbooks Gold Author and watsonx.ai Practitioner Mastery L5.

Senior AI Engineer

• Scuderia Ferrari HP × IBM: AI architecture for the Ferrari fan app. Ingestion of around a million data points per car per race and a multi-agent system that writes personalised race insights within a couple of hours of the chequered flag, with guardrails and quality metrics for IBM and Ferrari reviewers. Fan engagement grew 30% in the first year.
• Cultural heritage digitisation: proposed and won 12 AI use cases for a national programme (ancient texts, historical maps, audio, video, images, musical scores), then owned the hybrid on-premise and Azure architecture, scaling to more than 10M samples a month on GPU clusters.
• Consiglio di Stato: defined the legal-document anonymisation pipeline (LLMs, fine-tuned NER, deterministic rules) and coached the client's data science team to run it independently.
• Speaker at IBM events for government and public-sector clients.

AI Engineer

• IBM Research Zurich: first author of Zero-Shot-BERT-Adapters, a multilingual pipeline for discovering unseen user intents, published in Findings of EMNLP 2023.
• Production conversational AI for telco: Watson Assistant, speech-to-text and text-to-speech, semantic clustering, NER and automated call testing in CI/CD.
• Python automation for fuel-processing reporting in energy and utilities.

Master Thesis Researcher

HErBERT: privacy-preserving NLP for text classification, running transformer inference on homomorphically encrypted data. The starting point of my later patent work on FHE-secured AI.

Computer Science Book Reviewer

Invited by Pearson to read and publicly review computer science books, mostly on programming languages.

Software Engineer (remote)

Adapted a Java Pong game I had published for Oracle's JavaOne event, where it ran on a big screen with vintage controllers for attendees to play.

Software Engineer

Built an Android messaging app and its Java server on a client-server architecture.

Zero-Shot-BERT-Adapters: a Zero-Shot Pipeline for Unknown Intent Detection

A two-stage method for multilingual intent discovery: a transformer fine-tuned with adapters for natural language inference, then used to find intents it has never seen, in English, Italian and German.

Patents

• FHE-secured enhanced transformer-based AI module, US 19/055,180
• Cryptographic watermarking for large language models
• Further filings in trustworthy AI and agentic security

Simplify Your AI Journey: Unleashing the Power of AI with IBM watsonx.ai

Simplify Your AI Journey: Hybrid, Open Data Lakehouse with IBM watsonx.data

Enriching Linux on IBM Z Workloads with AI

danielx

A long-running series I host to make AI and generative AI topics approachable.

M.Sc. Computer Science and Engineering

Specialisation in Artificial Intelligence and Machine Learning. Thesis on privacy-preserving transformer inference with homomorphic encryption.

B.Sc. Computer Science and Engineering

Computing systems engineering, with projects spanning file systems, compilers, FPGA hardware design and networked multiplayer software.

Information Technology Diploma

IBM Redbooks Gold Author

watsonx.ai Practitioner Mastery, Level 5

Fundamentals of Accelerated Computing with CUDA C/C++

Machine Learning Specialist, Advanced

Applied Machine Learning in Python

CCNA Routing & Switching