CareMind drafts Thai and English nursing notes with source citations, using a locally hosted 8B Thai medical LLM. I owned model serving, evaluation and AI operations (MLOps): running the model reliably on an on-premise GPU server, measuring whether its output is accurate, and keeping the service healthy.
- 01Patient recordsSource data for the note
- 02LLM drafts the note8B Thai medical LLM via Ollama
- 03 · built by meDrug-name repairRestores correct drug names
- 04Source citations97.7% citation precision
- 05Nurse reviewNurse checks and approves the draft
Serving and deployment
- Deployed the LLM with Ollama behind a secure, authenticated API on an on-premise GPU server
- Containerised the backend with an automated GitHub Actions image workflow and a one-step Docker Compose setup
Evaluation and testing
- Built automated tests that run in CI before every merge, and measured 97.7% citation precision
- Compared Q3 and Q4 quantisation over 42 runs each, found a significant drug-name regression in Q4 (p = 0.029), and built the repair step that fixes drug names before citations are attached
Reliability and operations
- Load-tested 3,806 requests over 3 hours with zero errors or slowdown, and no hangs under burst load
- Set up GPU monitoring: a self-healing watchdog, alerts, real-generation uptime checks, daily accuracy spot-checks and backups
- Diagnosed a GPU memory lock-up and automated its detection and release


