AI/ML Engineer · LLM Systems · GPU & HPC — 📍 Padova, Italy · open to relocation / remote

I build and ship LLM and ML systems end-to-end — from production RAG pipelines and LoRA-fine-tuned small models to GPU-accelerated workflows running on consumer and HPC hardware. Ph.D. in Computer Engineering, 4+ years bridging applied research and production, with industry experience at MAIOR SRL (European leader in public-transport software) and open-source work used by teams at Harvard, NYU, and CUNY.

Looking for applied research / research engineer roles on LLMs, agents, or AI for software.

• MARQ — Multi-Agent Reasoning for Requirements Quality — Multi-agent system (Scanner → Investigator → Critic → Synthesizer) that detects ambiguity defects in industrial software requirements from text alone, using five small open LMs (Qwen2.5, Llama 3.2, Gemma 2, Phi-3.5) without fine-tuning. Reformulates QuRE to the harder requirement-level setting (no weak-word hints at test time), benchmarks 4 methods × 5 models × 5 folds, and uses the agents' reasoning traces to audit QuRE's gold labels via a blind human study. Target: ICSE 2027 NIER.
• QuRE LoRA Lab — LoRA fine-tuning of open small LMs (Qwen2.5, Gemma 2, Llama 3.2) for ambiguous-language detection on 2,187 real industrial software requirements. Gradio app, OOD evaluation, no API dependency.
• CudaMon — Open-source R package wrapping NVIDIA NVML for GPU telemetry in long-running scientific workflows. Used in CZI-funded single-cell genomics pipelines (Harvard, NYU, CUNY) on the Bioconductor ecosystem (>10M annual installs).
• Talk-to-PDF (MAIOR SRL, 2023–2024) — Production RAG chatbot over technical PDFs with custom chunking/embedding pipelines. Dockerized, deployed into the company's product stack, used internally by transport-planning engineers.
• AI Trip Classification (MAIOR SRL, 2022–2024) — ML pipelines for labeling planned trips from noisy GPS/AVL data; scheduling optimized with Resource-Constrained Shortest Path solvers. Deployed across European transit operators.
• Automatic Prompt Engineering — Backtracking-enhanced APE for LLM-based classification. CLI + Gradio replication package. REFSQ 2025.
• LLM Indoor Navigation (CNR-ISTI, 2025) — Multimodal LLM that interprets indoor maps and gives turn-by-turn guidance.

Languages Python · R · C++ · SQL · Bash ML / LLM PyTorch · Hugging Face Transformers & PEFT · LoRA / QLoRA · vLLM · Ollama · LangChain · LlamaIndex · sentence-transformers RAG & Agents FAISS · Chroma · retrieval pipelines · agentic architectures GPU / HPC CUDA · NVML · GPU profiling · CUDA-aware containers Serving & Ops Docker · FastAPI · Git/GitHub · CI/CD · MLflow · Weights & Biases Data Pandas · NumPy · Polars · PostgreSQL · Bioconductor

• Shipped LLM products inside a European transport-software vendor (MAIOR SRL)
• Open-source GPU tooling adopted by NIH- and CZI-funded genomics consortia (Harvard, NYU, CUNY)
• 70+ citations · h-index 4 · first author at REFSQ, CIBB, QUATIC · Q1 journal publication
• Program Committee: AIRE 2026 · Prompt-SE 2026

· ·