Study describes deployment of the LLENS platform for AI agents within a closed hospital network
The study authors deployed the LLENS platform within the closed network of Hokkaido University Hospital. It runs on eight NVIDIA H200 GPUs with the Kimi K2.6 model. When working with hospital documents, they measured accuracy of 97.2 % with an interactive approach compared with 81.3 % with fixed retrieval.
The study authors developed and deployed the LLENS platform, which, according to their report, has been running since 14 September 2026 within the closed network of Hokkaido University Hospital. It uses a single server with eight NVIDIA H200 GPUs, the open-weight Kimi K2.6 model with approximately a trillion parameters, and a web application rebuilt from Open WebUI. The authors measured a speed of 119 tokens/s for a single user and an aggregate speed of 908 tokens/s at 16 concurrent requests.
The architecture includes reading page images, full-text attachments, a document tree addressed by paths without a vector index, a network-isolated sandbox, checks of generated outputs, subagents, and saved workflows. On 180 questions based on hospital documents, the authors measured average accuracy of 10.7 % without documents, 81.3 % with fixed retrieval of ten text passages, and 97.2 % with an interactive approach using document tools and the sandbox environment. The difference was 15.9 percentage points, with a 95% confidence interval of 10.9–21.3 points. For questions requiring filtering and counting, the fixed approach achieved 16.7 % and the interactive approach 87.8 %.
Using synthetic data, the authors tested five scenarios: a procedure following a needlestick injury, batch assessment of adverse event severity, transcription of a scanned referral letter, aggregation and visualization of falls data, and investigation of three topics using subagents. These demonstrations establish the feasibility of the described workflows, but do not by themselves confirm their reliability when working with real patient data.
Why it matters
The study provides a concrete architecture for hospitals whose information systems operate within a closed network. The measured results indicate that the ability to repeatedly search, read, and perform calculations may be important when designing document workflows, especially for tasks requiring filtering and counting.
Two audiences, two different impacts
What this means
For individuals
The study shows physicians potential uses for local agents in transcribing referral letters or processing adverse events. These workflows were tested with synthetic data, so their benefits for everyday care remain uncertain.
More practical updates →For a business
Hospitals gain a concrete reference for designing an internal platform: the deployment described runs on a server with eight NVIDIA H200 GPUs and includes performance measurements at 16 concurrent requests. This helps assess the infrastructure and capacity needed for similar operations.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.