You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$1,500 left to reach the funding goal.
Project summary
My research is on AI in clinical decision support, mainly on how we test language models before they give health advice to patients. I studied medicine, then moved to AI engineering, and now I’m working in multidisciplinary research as a postgraduate student. When I was working in clinic I saw what happens when triage goes wrong, so this topic is personal for me.
For the first time, I got two papers accepted at NeurIPS 2026 in Sydney. The first one is at the GenAI4Health workshop (12 Dec) and the second at the MuslML workshop (7 Dec). This work has no lab or grant funding.
Papers:
GenAI4Health: https://neurips.cc/virtual/2026/loc/sydney/177660
What are this project’s goals? How will you achieve them?
In the GenAI4Health paper I look at how medical chatbots are tested. Most studies give the model the full patient case in one message and then score the answer. Real patients don’t do that. They forget things or don’t know which detail matters, so the doctor has to ask, and a safe model should ask too. In one study a system made zero mistakes in single-turn tests but had 40 to 50% errors in multi-turn conversations. I suggest three things that benchmarks can add. The first checks if the model asks the questions the case needs without asking too many. The second compares the same case with and without the key detail. The third follows how safety changes as the conversation gets longer.
The MuslML paper asks if a model that is safe in one clinical area is also safe in another, and the studies I reviewed show it is not. In emergency cases the models made mistakes at both ends, and one stress test missed 51.6% of real emergencies. In suicide risk the mistakes were mostly with middle-risk patients. So a test built only on emergency cases will not find the suicide-risk problem. I also found that when a model was grounded in a nurse triage protocol, its sensitivity for real emergencies went from 0.73 to 0.98, which was more than the difference between any two models.
In Sydney I want to present both posters and get honest feedback from people who build evaluations. I also want to find people to work with me on the next step, which is a set of triage conversations written and checked by doctors.
How will this funding be used?
Only for the flight. A Qatar Airways return ticket from Cairo to Sydney for 5 to 13 Dec is about $1,500. $1,000 would cover most of it. NeurIPS is already paying my registration and hotel (6 to 13 Dec).
What are the most likely causes and outcomes if this project fails?
If the project doesn’t reach the minimum, nobody will be charged. I won’t be able to present, because presenting is in person only. I will ask NeurIPS to release my spot so someone else can use it.
How much money have you raised in the last 12 months, and from where?
NeurIPS pays the hotel and registration directly. For the flight I applied to BlueDot Rapid Grant and Emergent Ventures, and both rejected me. I asked my university as well but they couldn't support.