Personal project · Applied AI
Health Insights turns WHO health data into a conversation, giving grounded answers to the people who set policy.
About
The data that should drive health policy usually sits in files nobody has time to read.
Effective public health decision-making depends on the timely and accurate analysis of vast and complex health data. Policymakers, healthcare administrators, NGO directors and researchers often struggle to access and interpret this data efficiently, which leads to misinformed decisions. Health Insights addresses this with a domain-specific chatbot: stakeholders query WHO health indicators in natural language and receive context-aware, accurate insights without needing technical expertise.
How it works
Ask
A stakeholder asks a question in natural language. No dashboards and no query languages.
Retrieve
A Retrieval-Augmented Generation pipeline retrieves the relevant WHO health indicators as grounding context.
Answer
A fine-tuned LLaMA-2 model responds with a context-aware, accurate insight.
Under the hood
General-purpose language models guess, and in public health a guess is a liability. The system therefore grounds every answer in retrieved data. A Retrieval-Augmented Generation pipeline finds the relevant records first, and the model only reasons over what it was handed.
The model itself is LLaMA-2, fine-tuned with QLoRA, a parameter-efficient technique that adapts the model to the health domain on modest hardware. Task-based user studies showed improved efficiency, clarity and confidence compared to traditional manual methods.
Mean rating by dimension
percentage of maximum score
What evaluators agreed with
percentage of maximum score
- 90%
- would use it again · 4.5/5
- 88%
- overall satisfaction · 4.4/5
- 100%
- task success rate
- 15
- evaluators, 5 task types