# SpiralAI

- **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:** [Danica Catoni](https://cerebralvalley.ai/u/danicat), [Mohammad Alkhatab](https://cerebralvalley.ai/u/mkhatab)
- **GitHub:** https://github.com/khatabx-king/spiralai
- **Demo video:** https://youtu.be/B3skPY7vs0w
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/796

SpiralAI is a physics-verified scientific design agent that uses Meta Muse Spark Open to explore theoretical graphene mechanical energy-storage designs.

Muse serves as the scientific-design reasoning component. It proposes candidate strain values, interprets verified iteration history, and adapts subsequent proposals based on feedback. Muse does not calculate or certify the numerical physics itself. Each candidate is passed to a deterministic Python physics engine, followed by an independent verifier that determines whether the calculated evidence supports the target, rejects the candidate, or indicates that further validation is required.

The core closed-loop workflow is:

Muse proposal → deterministic physics → independent verification → verified feedback → redesign or stop

For one recorded 50 Wh/kg Muse trajectory, Muse proposed 5.0%, 7.0%, and 7.1% strain. The deterministic physics engine calculated approximately 26.07, 49.61, and 50.96 Wh/kg, respectively. The independent verifier returned TARGET_REACHED on the third iteration.

SpiralAI also includes a deterministic optimizer that independently searches for the minimum strain required to satisfy a target. For the 50 Wh/kg benchmark, it identified approximately 7.03% strain, corresponding to approximately 50.03 Wh/kg in the current theoretical model. This provides a deterministic reference against which AI-generated candidates can be evaluated.

For a 100 Wh/kg target, the current 10% analytical screening boundary is insufficient. SpiralAI therefore does not claim a validated solution within the screening range and instead identifies the case as requiring further validation.

This separation between AI reasoning, deterministic calculation, optimization, and verification is deliberate. Muse explores the design space and adapts using verified evidence, while deterministic tools remain responsible for numerical scientific calculations and validation decisions. This prevents unsupported AI-generated numerical claims from becoming accepted scientific results.

The project includes a Streamlit Agent Console with Live Muse, Deterministic Optimization, and Provenance Replay modes. Append-only JSONL provenance records preserve recorded Muse runs, allowing previous trajectories to be inspected and replayed without requiring a live API connection. The repository also includes automated verification and currently passes 47 automated tests.

SpiralAI is designed as an extensible scientific-agent architecture. The current implementation intentionally uses a constrained graphene strain-screening problem as its initial test environment. The architecture separates reasoning, physics, optimization, verification, and provenance so that future versions can incorporate richer design variables and higher-fidelity scientific models.

All current SpiralAI results are classified as THEORETICAL_SIMULATION. Experimental validation has not been performed. The reported values are theoretical screening estimates and should not be interpreted as demonstrated device performance. Higher-fidelity modeling and experimental validation are required before practical performance claims can be made.

## More from Meta Global AI Developer Hackathon

- [DocTrust](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/793)
- [Ocular](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/794)
- [Solar Auditor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/795)
- [simTrain Universal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/797)
- [Rikky](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/798)
- [ctrl+v](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/799)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/796. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt