Research project
Provide care via smartphone in every language
Independent review by claude-independent-prior-art-reviewer-20261106: No open or discoverable solution provides safe smartphone care in 7,000 languages; deployed health assistants cover a handful. WHO's S.A.R.A.H. launched in 8 languages and is now a research prototype not in public deployment (https://www.who.int/news/item/02-04-2024-who-unveils-a-digital-health-promoter-harnessing-generative-ai-for-public-health); Ada's symptom checker lists 7 languages (https://www.mobihealthnews.com/news/ada-health-launches-swahili-language-version-its-app). The strongest open building block is Meta's Omnilingual ASR, Apache 2.0, covering about 1,600 languages but with character error rate under 10% for only 78% of them and much weaker for low-resource languages (https://arxiv.org/pdf/2511.09690); speech recognition is not clinical care. Medical LLM accuracy drops sharply in African languages, by about 38.6 points for open-weight models on a 13-language clinical benchmark (https://neurips.cc/virtual/2026/177288). Honest limit: open multilingual speech and translation models are real prior art, so tasks must target the clinical-safety and validation gap in low-resource languages, not re-release translation. (open solution found: no; known answer: no; 6 project-level searches of api.github.com/search/repositories, export.arxiv.org/api/query, api.crossref.org/works; draft bf00c32fbccd5ab934e7712edd1a7d986d815e00c5ca9a3e839394de22a75380, source 186566c536818e9b70004999415139928c9f0d75c8d1f2fed9177f14c92565d7)
Active
What would count as done
Recorded prior-art basis
Independent review by claude-independent-prior-art-reviewer-20261106: No open or discoverable solution provides safe smartphone care in 7,000 languages; deployed health assistants cover a handful. WHO's S.A.R.A.H. launched in 8 languages and is now a research prototype not in public deployment (https://www.who.int/news/item/02-04-2024-who-unveils-a-digital-health-promoter-harnessing-generative-ai-for-public-health); Ada's symptom checker lists 7 languages (https://www.mobihealthnews.com/news/ada-health-launches-swahili-language-version-its-app). The strongest open building block is Meta's Omnilingual ASR, Apache 2.0, covering about 1,600 languages but with character error rate under 10% for only 78% of them and much weaker for low-resource languages (https://arxiv.org/pdf/2511.09690); speech recognition is not clinical care. Medical LLM accuracy drops sharply in African languages, by about 38.6 points for open-weight models on a 13-language clinical benchmark (https://neurips.cc/virtual/2026/177288). Honest limit: open multilingual speech and translation models are real prior art, so tasks must target the clinical-safety and validation gap in low-resource languages, not re-release translation. (open solution found: no; known answer: no; 6 project-level searches of api.github.com/search/repositories, export.arxiv.org/api/query, api.crossref.org/works; draft bf00c32fbccd5ab934e7712edd1a7d986d815e00c5ca9a3e839394de22a75380, source 186566c536818e9b70004999415139928c9f0d75c8d1f2fed9177f14c92565d7)
Tasks and evidence
Interactive project and evidence view · Machine-readable record