AI Cloud
Building a Dementia Diagnosis LLM Chatbot with L40S GPUs and WebUI Templates | GDG Speaker Jeongju Lee
2025.09.13.
Turning an LLM project into a real, working service requires more than just running a model. Data preprocessing, training, artifact management, and web app integration all need to flow together as one stable development pipeline. Local PCs run into performance limits, and cloud environments like AWS require time-consuming initial setup—instance provisioning, port configuration, and access setup all add up. Jeongju Lee took on a project to build an LLM-based chatbot using dementia diagnosis data. The goal went beyond simply testing a model: he aimed to transform the data into a trainable format, fine-tune it with LoRA, and integrate the result into a Flutter web app as an actual service.
■ An LLM Development Environment That Started with a Single Template Selection
Jeongju Lee used Runyour AI to rent an NVIDIA L40S GPU for model training and chatbot testing. Depending on the task at hand, he alternated between the Ubuntu template and the Text Generation WebUI template. The Ubuntu template was a good fit for building custom models and training environments from the ground up, while the Text Generation WebUI template came preconfigured for chatbot testing—allowing him to verify results quickly through a GUI.
As a result, he could spend less time on environment setup and more time on model configuration and testing.
■ A Workflow Held Together by the Storage Feature
In LLM development, reliably managing artifacts—preprocessed data, LoRA training outputs, Python execution files, and more—is just as important as the modeling itself. Jeongju Lee used Runyour AI's storage feature to retain his data while switching between multiple templates. Files created in the Ubuntu environment could be reused in the WebUI template, and training outputs could be stored and carried over into the next stage of work. Because data persisted even when templates changed, the project flow was never interrupted mid-stream.
■ On-Demand GPUs That Reduced Setup Time and Cost Pressure
Colab, which he had previously used, was good for quick tests but limited for longer workloads. AWS came with the overhead of instance setup and cost management. And a local environment couldn't reliably handle larger models due to hardware limits. With Runyour AI, simply selecting a template and renting GPU resources on demand meant work could start almost immediately. Compared to AWS, Jeongju Lee found that setup time was cut by more than half, and costs were reduced by roughly 30%.
This project shows that even users new to LLMs can take on real, service-level work without wrestling with complex cloud setup. Pre-configured templates lowered the barrier of environment configuration, the L40S GPU provided stable support for model training and testing, and the storage feature created a development flow in which data and outputs carried over seamlessly from one stage to the next. Looking ahead, Jeongju Lee plans to use Runyour AI to build a travel companion chatbot trained on travel tips and trends. Runyour AI is well suited for newcomers who have studied LLMs through books or courses and now want to run real models, as well as for developers looking to connect web apps with AI models and turn the result into a working service.
