# Kitsa

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Rohit Banga](https://cerebralvalley.ai/u/kitsa)
- **GitHub:** https://github.com/kitsapublic/2026-10-muse-hackathon
- **Demo video:** https://youtu.be/lIPVKNjBGnM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4052

Kitsa — Simulate Clinical Trials

Kitsa builds AI agents for clinical research: they draft regulatory documents, find research sites and identify eligible patients. For this hackathon, we built Simulate Clinical Trials on Muse Spark, so researchers can test eligibility criteria against patient records while the protocol is still being written.

Trials run late because they cannot find enough eligible patients, and the cause is set before recruitment opens: the criteria that decide who can join are carried over from earlier protocols, in the lab, without looking at the patients hospitals actually see. Our application shows who qualifies, which criteria exclude people and what a proposed change does to the population before the protocol is final.

The demo uses an illustrative lung-cancer trial with ten criteria and 100 synthetic patient records. Muse reads the clinical notes, extracts the facts each criterion depends on, and cites the passage behind each one; the engine combines those facts with structured data such as age, lab results, and treatment dates. Waterfall charts show where the population narrows; patient-level views show the evidence behind each decision.

Under the draft criteria, 19 of 100 patients are eligible, and the age limit alone removes 26 before applying any clinical criterion. Raising the age limit from 75 to 80 gives 29; allowing three prior regimens instead of two gives 23; and both together give 37: almost double, from the same records, with the other eight criteria untouched. Patient 011, aged 78 with three prior regimens, qualifies only when both limits move. Thirty patients remain unresolved and are kept separate.

The proposed changes, population impact and evidence export for protocol review, where the clinical team decides what preserves safety and scientific validity. The live run screened 100 patients in about 66 seconds for roughly $0.65 of model usage; threshold changes need no new model calls, and the downloadable demo replays the genuine recorded responses without an API key.

The stakes are documented. Eligibility changes are the most common reason for substantial protocol amendments; each costs a median $535,000 on a Phase III trial and adds about three unplanned months (Tufts CSDD). At industry estimates of $600,000 to $8 million per day of delay, catching one eligibility problem before the protocol is locked is worth tens of millions of dollars per trial. In a Nature study of 61,094 lung-cancer patients, broadening restrictive criteria with real-world data more than doubled the eligible pool with little change in outcomes (Liu et al., 2021). Lung cancer alone affects about 2.5 million new patients a year. Designing criteria around real patient populations before the first patient enrols means faster trials, results that apply to more people who will take the medicine, and treatments that reach them sooner.

## More from Meta Global AI Developer Hackathon

- [Songbird](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4049)
- [Heavenly Technologies Open Source](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4050)
- [BonSafar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4051)
- [CLEVERS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4053)
- [Agentic Corporation](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4054)
- [Dreamwalk](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4055)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4052. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt