A healthcare assistant on an open model
95% answer accuracy and 80% shorter waits
- Sector
- Healthcare
- Market
- Singapore
95%answer accuracy, 30% better than before
80%shorter wait times
10,000+concurrent users
The challenge
A healthcare provider wanted a virtual assistant that answers patients from its own data, on a model it controls.
What was built
- A retrieval pipeline on Mistral-7B over the provider's own data
- Semantic search with re-ranking, and fact-checking of model output
- An indexing pipeline that turns the provider's records into searchable chunks
- Containerised API services, monitored with the ELK stack
How it was built
- An open-weight model, deployed in the provider's environment