# Radmesh

- **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:** [Miguel Angarita](https://cerebralvalley.ai/u/Angaritarad)
- **GitHub:** https://github.com/angaritamd/oncall-resident
- **Demo video:** https://youtu.be/kRR_YkbcYxU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1423

RAD·QC — On-call Resident

On a night shift, a radiologist reads dozens of studies. Two errors are common and preventable: a finding described in the body of the report but missing from the conclusion (the part referring doctors actually read), and a clinical question from the order, such as "rule out pneumothorax", that the report never answers. Both can delay treatment.

RAD·QC is a report checker fine-tuned on Meta's Muse Glimmer. It reviews every draft before signature, in about 1.4 seconds, and quotes the exact sentence it is concerned about. It is part of On-call Resident, a voice-first assistant that helps a radiologist run the shift from the phone while images stay on the workstation.

How it was built: Muse Spark Open generated 739 synthetic reports; I, a radiologist, reviewed samples from every batch and wrote the clinical rules (what counts as an omitted finding, which questions are critical, what is too minor to flag). 697 cases were kept and locked with a sha256 manifest, then used for a LoRA fine-tune on Muse Glimmer. Voice dictation uses Muse Voice Transcribe.

Results on our 66-case benchmark (mean of 3 runs): 62 correct vs 53.7 for base Muse Glimmer, about 10× faster (1.35 s vs 14 s median), and zero false alarms on correct reports.

Safety by design: synthetic reports and public DICOM only, a PHI scanner gates every commit, and only a human can sign. The AI never diagnoses and never signs.

Built solo by Miguel Angarita, MD, radiologist in Bogotá, Colombia.

## More from Meta Global AI Developer Hackathon

- [AccessiWebEngine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1420)
- [PayGuard](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1421)
- [TallyHealth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1422)
- [Rastro](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1424)
- [Eternity Nest](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1425)
- [Abhi-the-great](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1426)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1423. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt