A complete system deserves a complete partner.

A prioritised map of use cases; an estimate of users and load (tokens per day, peaks, concurrency); a reasoned choice of main and specialist models, with verified license; recommended and alternative system size; a site energy balance with month-by-month solar generation, night-time autonomy and reserve; a first GDPR and AI Act assessment; an economic estimate.

A meeting with the people who will use the system, collection of sample documents, an on-site or remote survey, a written report with stated assumptions. If it is not worthwhile, we say so. The cost is deducted from the order.

PVGIS for solar generation, public model benchmarks, ML.ENERGY for energy per response.

A trial environment with the chosen model; the customer’s real document archive made searchable; a set of 50–100 real questions with expected answers, prepared together with the customer; a measure of quality, speed and consumption; a final report with the outcome.

Data handled under a confidentiality agreement and appointment as data processor; documented deletion at the end.

lm-evaluation-harness and a custom evaluation set in Italian or the customer’s language.

Assembly and wiring in Padua; 72 hours of operation under load; a battery switchover test with no interruption for the GPUs; a cooling-failure test; models pre-loaded and verified; a signed acceptance record.

A test protocol shared in advance; the customer may attend.

JA Planet BESS standards (IEC 62619, IEC 62933-5-2, UL 9540A), EN 50600 as a reference for the compute-module infrastructure.

Supervision of the local installer; energy and compute commissioning; site acceptance test (SAT); administrator training (user management, updates, backups) and role-based user training; a manual in Italian or the customer’s language.

CEI 0-16 / CEI 0-21 connection standards where the grid is present; the site’s electrical safety procedures.

Archive import (PDF, scans, email, ERP) with OCR; meaning-based search with source citation in every answer; the customer’s terminology, acronyms and language through light model adaptation (LoRA); roles and permissions aligned with the organisation; integrations with the clinical record (HL7 FHIR, DICOM for images), records management, ERP and email.

Short cycles, each one verified with users against the pilot question set.

Docling and PaddleOCR-VL for documents, Qdrant for the vector archive, vLLM for serving.

Data protection impact assessment (DPIA, Art. 35 GDPR); update of the processing register; documented security measures (Art. 32 GDPR); AI Act documentation: user notice (Art. 50), use logs and human oversight for high-risk uses (Art. 26), support for the fundamental-rights impact assessment for public authorities (FRIA, Art. 27); model sheets with license and limits.

GDPR, AI Act (EU Reg. 2024/1689 as amended by the Digital Omnibus), ISO/IEC 42001 for the AI management system, ISO/IEC 27001, NIS2 directive for in-scope entities. Note: technical-documentary support, not legal advice.

Continuous monitoring of energy (generation, state of charge, auxiliaries) and compute (GPUs, temperatures, queues, response times); alerts; scheduled security patches; encrypted backups; a monthly report with availability, energy used and kWh per million tokens.

A separate, encrypted management channel; no user data leaves the site.

A quarterly assessment note; an update tested on the customer’s question set.