Healdar / Independent product · Generative AI
Making regulatory answers traceable.
A bilingual AI assistant for navigating health-AI regulation across the Gulf, Europe, the United States, and international guidance.
- My role
- Independent product builder; concept through deployment
- Period
- 2026 · Version 2.1 released 14 September
- Status
- Public beta · Open source
English / العربية
Ask a question.
Follow the source.
A practical application of retrieval-augmented generation.
The product problem
Health-AI regulation spans different jurisdictions and long documents. Finding a relevant passage, comparing requirements, and working across English and Arabic can make even an initial research task time-consuming.
Healdar applies retrieval-augmented generation to that navigation problem. It brings an answer together with references to the source document and page, so the person asking can inspect the supporting material.
From concept to public beta
I built and deployed Healdar independently, covering the product concept, bilingual experience, document processing, retrieval pipeline, and application. My prior work on medical software regulation helped define the use case.
Version 2.1 is grounded in 63 official documents across Saudi Arabia, the UAE, Qatar, the EU, the US, and international bodies including WHO and IMDRF. Users can ask in English or Arabic, describe a product scenario, continue with follow-up questions, and compare jurisdictions side by side.
Making the answer useful
Make the source visible.
Answers link to source documents and pages, with passages available to inspect. Citation numbering stays consistent across the screen, PDF, and Word exports.
Support a line of inquiry.
Follow-up questions retain the conversation’s context and supporting pages. Side-by-side comparisons help users investigate a question across two jurisdictions.
Work across languages.
Multilingual retrieval connects English and Arabic questions with source material in either language, including official Arabic documents.
Make limitations visible.
A relevance gate declines questions without suitable retrieved evidence. Coverage indicators flag when the available documents only partly address a question.
A focused RAG implementation
The pipeline combines multilingual embeddings with BM25 keyword search, so retrieval can match both a concept and a specific article or document identifier. Query planning helps find provisions relevant to a product scenario, while jurisdiction balancing prevents a larger document collection from dominating cross-jurisdiction questions.
The implementation uses Python, Streamlit, ChromaDB, and GPT-OSS models through Groq. Automated tests and a curated evaluation set check retrieval, refusal behavior, citations, and exports.
Healdar is a navigation and research assistant. Regulatory conclusions still need to be checked against the authoritative documents and the product’s specific context.
Try it or explore the code
The public beta and version 2.1 source code are available to explore.
Next project