AI / ML Engineer · Biomedical Engineering · MLOps
I build machine learning that has to survive contact with real clinical data.
Biomedical engineer turned AI/ML engineer, with six years across healthcare machine learning, MLOps and cloud-native systems — enterprise consulting at PwC and Capgemini, biotech instrumentation at Tecniplast, and applied research at Fondazione Bruno Kessler and Deep Learning Italia, with three peer-reviewed publications on MLOps and biomedical imaging.
Currently Co-Founder & CTO at Akaion, an AI-native orchestration platform I architected and scaled to 1,250+ users and multiple B2B clients in under a year. Our sovereign execution kernel, Annona, is open source.
I hold an MIT Sloan MicroMasters in AI in Healthcare and completed ETH Zurich's Machine Learning in Health programme. Based in Como, Italy — twenty minutes from the Swiss border.
Where an AI workload runs is a decision — and it should be yours, enforced and recorded.
Annona is an execution kernel that decides where each step of an agent runs (local, on-premise, or cloud), enforces that decision with a default-deny policy engine, and writes every placement to a hash-chained audit ledger. Local by default, GDPR by design, aligned to the EU AI Act. Apache-2.0, 559 tests, Docker and pip install annona.
It exists because regulated industries — healthcare above all — cannot send data to whichever endpoint a model happens to prefer, and cannot deploy what they cannot audit.
Read the design · Docs · Repository
Languages
AI & Machine Learning
Cloud & MLOps
Backend & Data
Frontend & Devices
| Project | What it does | Stack |
|---|---|---|
| MeHEDI-app | Data-driven management platform for a small healthcare facility — full-stack, from data model to dashboard | Python · Streamlit |
| medsync_hl7-FHIR | Clinical data interoperability with HL7 and FHIR — the layer that makes health data actually usable | Python · FHIR |
| mlops-api-cell-counting | Cell counting served as an API, with the MLOps discipline around it: reproducibility, versioning, monitoring | Python · FastAPI · Docker |
| API-Ultrasound-Classificator | Multi-class classification of ultrasound images, exposed as a production API | Python · Deep learning |
| MLODC-Transfer_Learning | Multi-label classification in ophthalmology via transfer learning | Keras · TensorFlow |
| go_rest_biomed_tracking | REST service for biomedical tracking, written in Go | Go · REST |
| Neuroimaging-Thesis-Project | 3D U-Net benchmarks for paediatric brain MRI segmentation — the work behind my thesis and a publication | Python · PyTorch |
- M. Testi, M. Ballabio et al. (2025). MLOps: A Use Case in Biomedical Image Classification. Medical & Biological Engineering & Computing.
- M. Testi, M. Ballabio et al. (2022). MLOps: A Taxonomy and Methodology. IEEE Access.
- Amorosino et al. (2021). DBB — Distorted Brain Benchmark. Brainlife.io (contributor).
