# Alex Sumaila

- **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:** [Alex Sumaila](https://cerebralvalley.ai/u/alexsirguy)
- **GitHub:** https://github.com/Alexsirguy/SBI
- **Demo video:** https://youtu.be/Qvz5s02QVJA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3581

Everything around us runs on science, but almost nobody explains the thing right in front of you. Why is the bike easier to pedal uphill in a lower gear? Why does a kettle click off at the right moment? You can search for it, but you get a wall of links or a twenty-minute video that never quite answers your question.

Science Behind It turns whatever you're curious about into a short lesson you can actually play with. Snap a photo of an object, or just type a sentence like "bicycle gears". We recognise what it is and you ask your own question. A lesson is then written fresh, just for that question. There's no catalogue, so it works for almost anything, in English or Spanish.

Each lesson follows Merrill's First Principles of Instruction. It starts with a real problem and asks for your guess before explaining anything. At its centre is an experiment built for that object: a 3D mechanism you can pull apart, a molecule you can turn, or a graph you can change. If you're still curious, the tutor answers in a couple of sentences from the lesson itself. It can also reach into the experiment and show you. Ask what happens with a bigger rear sprocket, and the gears on screen change to match.

I cared a lot about trust:
- Privacy: faces in your photo are blurred on your phone before anything is uploaded.
- Honesty: when part of a lesson comes from the model's general knowledge rather than a source, the lesson says so plainly.
- Safety: answers to off-lesson questions are labelled as unchecked, and risky questions get a safe answer. Safety combines keyword rules with an AI classifier that can raise a risk level but never lower it.

Under the hood:
- The codebase: a TypeScript monorepo of 18 packages behind a NestJS API, with 25 written architecture decisions.
- Visuals: the model writes a spec (MuJoCo, PubChem, Mermaid or Vega-Lite), and we check it before drawing it. Only when nothing fits does it write code, and that code runs in a sandbox with no network access.
- Testing: we measure all of this with visual and tutor evals, weekly lesson benchmarks, and performance budgets enforced in CI.

I want anyone, whether a curious kid or an adult who never "got" physics, to be able to point their camera at anything and find the science behind it.

## More from Meta Global AI Developer Hackathon

- [OmniSpatial](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3578)
- [Persona Lab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3579)
- [FaulltLine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3580)
- [newcase.ai](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3582)
- [SIAGA.ID](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3583)
- [Mihail Cernei](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3584)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3581. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt