Authors - Andrei-Horia Ignat, Dan-Matei Popovici Abstract - Large language models solve tasks across many domains, but their size and reliance on hosted providers place them out of reach of organisations that must own their deployment. We present a multi-model, multi-adapter framework for building and serving a conversational multiexpert search platform entirely on commodity hardware. Each capability is realised as a small, task-specialised expert: a parameter-efficient (QLoRA) adapter over one of a few shared base models, trained on data synthesised offline by a large model rather than collected from users. A lightweight router directs each query to the relevant expert, which answers by emitting an executable tool call against the service’s data. Because every expert is a few-megabyte adapter over a shared base, the dominant memory cost is the base models, not the number of experts: a single workstation GPU holds an estimated several hundred to several thousand experts, and adding a service requires only a short, repeatable recipe rather than an advanced MLOps stack. We organise the design around eight requirements spanning commodity-hardware operation, extensibility, accessibility, multilingual access, data-grounded answers, transparency, self-hosting, and graceful integration depth, and we describe serving optimisations (multi-adapter co-residency, prefix caching, function-call routing) that keep it efficient. We instantiate the framework on the FUTURAL Metasearch Platform, the smart-solution registry of an EU agricultural project; a representative wildlife expert reaches about 95.1% token-level accuracy after roughly twenty minutes of training on one GPU.
