# SPECTRA AI Research

- **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:** [Sergo Episkoposian](https://cerebralvalley.ai/u/Sergoep)
- **GitHub:** https://github.com/sergoep/SPECTRA-ECG
- **Demo video:** https://youtu.be/W5nrUFBj-Og
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/785

SPECTRA-ECG is a non-diagnostic research prototype for deciding where expensive multimodal AI attention should be spent in long ECG recordings.

The system combines a frozen spectral scout with Meta Muse Spark 1.3 through the Meta Model API. Instead of sending every ECG window to the multimodal model, SPECTRA first ranks the signal using lightweight spectral features and routes only a fixed-budget subset of 5-second lead-II windows to Meta. Muse then produces structured review evidence for the routed windows. It does not receive the hidden endpoint labels or the spectral scout score.

Our locked hackathon evaluation contains 1,440 ECG windows from four MIT-BIH records with 2,045 endpoint events. At a 10% routing budget, SPECTRA sends 144 windows instead of 1,440, reducing logical Meta token use by 90.02% and logical route cost by 90.06%, while retaining 11.74% of FULL end-to-end endpoint recall. At a 20% budget, it sends 288 windows, reducing logical token use by 79.91% and route cost by 79.78%, while retaining 22.20% of FULL end-to-end recall.

The contribution is not a claim of improved medical diagnosis. SPECTRA demonstrates a resource-aware architecture in which inexpensive mathematical signal analysis determines where expensive multimodal AI computation should be allocated.

The public repository includes the locked experiment, results, figures, reproducibility instructions, independent verification scripts, API workload accounting, and audit materials.

## More from Meta Global AI Developer Hackathon

- [Torqdrop](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/782)
- [PocketDevs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/783)
- [EduSphere Safety](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/784)
- [VidMind](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/786)
- [SoloSmith](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/787)
- [Lucid](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/788)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/785. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt