The problem
Global clinical tools answer from US and UK guidelines. General chatbots answer confidently, often without a source. A doctor cannot use an answer they cannot check.
The rules, written before any code
- No uncited claims. Every statement links to its source.
- Refuse rather than guess.
- Use only retrieved sources, Indian guidelines and medicines first.
- A reference tool, not a diagnostic device. No patient data.
The release gate
Nobody edits a reference answer to make a test pass. When a doctor flags a wrong answer, it becomes a new test case.
What I left out
A custom or fine-tuned model. The lasting advantage is the Indian clinical corpus and citation accuracy.
Where it is
The evaluation sets and most of the pipeline are built. The gate has not yet run end to end on real answers.