I build AI that runs where the data lives — quantised transformers on microcontroller NPUs, self-hosted LLM and RAG stacks, and vision-language agents that never touch a third-party API.
Fig. 01 — Profile
On-premises deployment, secure execution and documentation discipline — plus a decade of translating dense technical material for non-technical audiences.
M.Sc. Computer Science & Engineering student (AI track) at Politecnico di Milano, with hands-on experience deploying open-source AI on local Linux infrastructure under strict data-sovereignty and hardware constraints — from Transformers running locally on STM32 NPUs, to Retrieval-Augmented Generation pipelines over 7,600+ document knowledge bases, to vision-language document-classification workflows. All on local hardware, minimising third-party API dependencies.
Focused on reproducible AI systems: secure execution, confidentiality, and systematic documentation of every experiment and configuration — version-controlled Conda environments, logged runs, nothing tribal-knowledge-only. Daily Ubuntu Linux user.
Alongside the engineering, several years spent making complex technical content land with mixed audiences: as a certified industrial-robotics trainer for FANUC and COMAU platforms, and as a guide at Milan's Leonardo da Vinci Science Museum. It's the same skill either way — knowing what to cut.
Fig. 02 — Toolbox
Grouped the way I actually reach for them, not alphabetised for show.
Fig. 03 — Selected work
Every number below is from the repo's own eval run — not a claim.
Designed, trained and INT8-quantised a 4.56M-parameter encoder-only transformer from scratch, then deployed it end-to-end on a microcontroller's Neural Processing Unit.
73.14 ms/token · 257/407 ops NPU-accelerated View on GitHub ↗Re-evaluated and extended the Sen2Fire benchmark with an nnU-Net baseline across the full Sentinel-2/5P spectral range, plus a test-time-augmentation cost/benefit study.
F1 0.3675 — new benchmark best (vs. 0.2810) View on GitHub ↗Swapped GPT-4o for locally-hosted Qwen3-VL, built a zero-manual-annotation classifier for a 1,101-image dataset, and fine-tuned a ResNet-18 visual prior into the agent loop.
62.7% task success — above the original paper's 59.1% View on GitHub ↗Benchmarked three single-hop retrievers and two LLM-driven multi-hop pipelines over a 7,600+ snippet knowledge base, served fully on-prem via 4-bit quantised CodeLlama-7B.
+38% relative CodeBLEU over the no-retrieval baseline View on GitHub ↗