# SciSleuth AI

- **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:** [mustapha Chaibi](https://cerebralvalley.ai/u/mustapha-chaibi)
- **GitHub:** https://github.com/chaibi-mustapha/scisleuth-ai
- **Demo video:** https://youtu.be/t1JJjIXsIvI
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2786

As an independent researcher and developer, I built SciSleuth AI after watching scientific controversies like LK-99 unfold online. I realized that scientific breakthroughs rarely stall because of a lack of intelligence—they stall because critical evidence is scattered across incompatible silos: conflicting lab protocols, unread figure telemetry, and paper limitations that another lab already solved years ago. A single researcher simply cannot cross-examine hundreds of multi-lab parameters in their head.

To fix this, I created SciSleuth AI: an autonomous "cold case" investigation operating system powered by Meta's Muse Spark 1.3 model. 

Instead of building another generic chat interface, I wanted a cinematic, tangible Evidence Board with real-time catenary physics strings. Here is how I structured the investigation:

• 6 Specialized Detective Agents: I built a sequential reasoning chain where each agent feeds the next. It starts with corpus harvesting and metadata extraction, decodes complex visual charts and operando microscopy videos, hunts down protocol contradictions (red strings), matches Lab A's limitations with Lab B's methods (green strings), stress-tests findings through an adversarial critic, and finally synthesizes a unified 60-day inter-laboratory action warrant.

• Real Figure & Video Telemetry: Rather than treating papers as plain text, I gave the system vision capabilities to decode actual micrographs, resistance curves, and operando films with automated timestamp navigation (like jumping to dendrite nucleation at 00:08).

• Live Dispute Arbitration: When two world-class laboratories disagree, I programmed the system to adjudicate the protocol clash live, evaluate confidence, propose a decisive test, and turn the contradictory string gold.

• Custom Dossier Builder: Alongside the 4 pre-loaded cold cases (LK-99, Alzheimer's tau protein, Lithium dendrites, CRISPR off-targets), I built a custom builder that queries arXiv and PubMed in parallel, extracts text from custom uploaded PDFs, and automatically drafts balanced thesis/antithesis dossiers in seconds.

I engineered the entire stack to be lightweight and rock-solid: FastAPI on the backend with smart 10/h rate-limiting and graceful offline fallbacks, paired with vanilla modern JavaScript and SVG for 60 FPS physics without heavy front-end dependencies. 

I hope SciSleuth AI demonstrates how modern multi-agent reasoning can turn fragmented science into closed cases.

## More from Meta Global AI Developer Hackathon

- [Arnav.S*k-3](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2783)
- [Zizo's Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2784)
- [Vist](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2785)
- [Tiny Libraries](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2787)
- [Kernelworks](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2788)
- [grinflow](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2789)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2786. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt