Building a sovereign clinical AI without an IT department: cross-model arbitration, scope migration, and what a solo clinician-builder actually gets — an implementation report

Objective: To report the unassisted implementation of a fully local, retrieval-augmented clinical AI assistant by a practicing oral and maxillofacial surgeon with no formal information-technology background, in a language (Albanian) with no published LLM benchmark, and to report the implementation's outcomes honestly against its originally stated objective. Setting: A privately owned multi-operatory dental day hospital in Tirana, Albania. Sole builder and operator: the first author. Methods: Retrospective reconstruction from complete, timestamped exports of all three commercial large-language-model platforms used during the build (~7,500 messages, 19 April – 16 July 2026), combined with a controlled inference benchmark, cost accounting, operational usage measures extracted from the deployed system's logs, the dated repair of a retrieval defect, an operational incident log, and a post-deployment field probe. Objective persistence was measured by term-frequency analysis across the full corpus. Results: A locally hosted 31-billion-parameter model with retrieval-augmented generation, WhatsApp gateway and local speech-to-text reached working state three days after the first question was asked and remained in continuous production 92 days later. The only hardware bought for the project was a used 24 GB GPU (~EUR 366); it ran on a workstation acquired earlier for imaging, whose cumulative cost (~EUR 4,900) is attributed in full in the appendix. On day one of use of each platform, each gave erroneous guidance on the decision within its purview; the clinician rejected all three recommendations, correctly in each case. The operational objective that motivated the project — reducing attrition at front-desk capture and treatment-plan acceptance — was never measured, and its second half never mentioned again anywhere in the corpus. Over 92 days of continuous production the owner WhatsApp channel alone carried 1,215 messages across 41 active days — 135 dictated Albanian voice notes among them — all under a single user account (the first author's; staff tested in supervised sessions under it). The workload is recurring, not a one-off: the extraction day was the second-busiest on record (100 messages), including live Albanian clinical queries. A 53-day Albanian retrieval defect was repaired on day 84 (12 July) by a configuration change (§3.7, Appendix A2); Albanian knowledge-base replies have recurred as routine operation since, not as a single verification event. No comparative answer-quality score is reported; the construct is examined and declined. Conclusions: A solo clinician can build sovereign clinical AI at commodity cost within days. What he cannot do alone, this report suggests, is keep the clinical objective in view while doing so. The transferable contributions are a governance method and a cross-model arbitration discipline, both stack-agnostic and near-free.