Jua

Donated medical equipment needs ongoing maintenance, documentation, and access to technical support. When those resources are missing, hospitals can struggle to keep equipment in use. This gap was the starting point for Jua.

My research focused on the difficulty of finding service manuals and repair support, particularly for older equipment. Limited access to maintenance training and documentation can leave technicians without the resources they need. I wanted to explore whether software could make that information easier to find.

Jua explored how an AI-assisted interface could help hospital staff find troubleshooting information for donated medical equipment.

The prototype combined a chat interface for searching manuals with a directory for finding technicians by location and specialization.

I built Jua Chat around GPT-3.5 to let users ask questions about equipment manuals in plain language.

The document question-answering system retrieved relevant passages from manuals and used them to generate answers with source references. Those references gave users a way to check the underlying material.

The aim was to make technical information easier to navigate. The hackathon prototype demonstrated that workflow; it did not establish that the answers were reliable enough for equipment maintenance in practice.

Jua Connect explored a way to find people with experience working on specific equipment. The interface included filters for specialization and location.

The goal was to help hospital staff find in-person support when they needed expertise beyond what a document search could provide.

I prototyped the interface in Figma and built the frontend with Vue.js and Tailwind CSS. The backend used FastAPI, paper-qa, and custom prompts for document question answering.

I was pleased with how closely the finished prototype followed the design I had planned, despite having only a few days and limited time to work on it.

Building the backend also gave me practical experience with large language models and natural language processing, particularly retrieving information from documents and using it in generated answers.

Time was the biggest constraint. My initial plan was ambitious for a short hackathon, and fitting the work into the deadline was difficult.

Integrating document question answering took more experimentation than I expected. I tried several models from Hugging Face, but their answers did not meet my expectations. I also considered LangChain before deciding I did not have enough time to explore it.

The research helped me understand how much equipment maintenance depends on access to documentation, training, and technical support. Those constraints shaped the problem I chose to work on.

On the technical side, I learned through experimenting with language models before choosing an existing document question-answering library. That let me spend more of the limited time connecting the backend to the interface.

After the hackathon, I identified answer quality as the first area to improve, including evaluating other text-generation models.

The prototype used web technologies and required internet access. I also considered a native app with offline capability to make it more useful where connectivity is limited.

Offline support would require working within device constraints. Running a model locally would need to balance answer quality with memory and processing requirements; using a cloud model would retain the dependency on internet access.

The short demo video submitted to CSYA Hacks 2023 shows the prototype in use.