# Bonisa

- **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:** [Sizanobuhle Zulu](https://cerebralvalley.ai/u/MAGEBA)
- **GitHub:** https://github.com/SizanobuhleZulu/BONISA
- **Demo video:** https://youtu.be/PTN0HFD7Utw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4195

Bonisa (“show” in isiZulu) turns a phone video of a physics experiment into a verified scientific measurement. We start with the pendulum, the classic lab for measuring gravitational acceleration g.

The problem: In most school and university labs, students time a pendulum with a stopwatch. Human reaction time (about 0.2 s) dominates the error. Results like 9.3 or 10.4 m/s² are common, and nothing tells the student whether the measurement can be trusted. Labs in under-resourced schools, like many in South Africa, can’t afford light gates or motion sensors, but every student has a phone.

How Bonisa works:

Meta SAM 3.1 (via the Meta Model API) tracks the pendulum bob through every frame from a plain-language prompt such as “orange ball”. It also finds an A4 sheet on the wall, which converts pixels to metres.
Physics engine (Python, NumPy/SciPy):
Fits a damped oscillation to the motion using each frame’s real timestamp, since phones record at variable frame rates.
Fits a circle to the arc to locate the pivot and measure the string length from the video itself.
Corrects for swing angle, then computes g = 4π²L/T² with a properly propagated uncertainty.
Self-verification: Eight physics checks:
tracking coverage
time resolution
number of swings
model fit
swing angle
camera tilt, detected from the A4 sheet’s aspect ratio
video length against the tape-measure length
final precision
Bonisa reports verified, verified with warnings or rejected. If any check fails, it refuses to give a number and explains in plain language how to fix the setup. An instrument that says “I don’t trust this” is more valuable than one that always answers.

Results so far:

On a simulated pendulum (true g = 9.81 m/s²), Bonisa measured g = 9.824 ± 0.042 m/s² and recovered the string length from video alone (69.9 cm vs 70 cm).
It correctly rejected a too-short recording and a tilted-camera recording.
The live SAM 3.1 integration is working end to end on real lab footage.
On a real chalkboard video where the bob was a tiny dark object, SAM found nothing, and Bonisa stopped instead of guessing.

Next:

Muse-powered explanations and re-filming advice in English, isiZulu and isiXhosa.
The same pipeline applied to springs, falling objects and projectiles, making Bonisa a general measurement layer for any lab with a phone.

## More from Meta Global AI Developer Hackathon

- [Triage](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4192)
- [Jay Xu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4193)
- [PostOrca](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4194)
- [Kaypro](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4196)
- [floodwatch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4197)
- [Verity — Nnaemeka Eziamaka](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4198)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4195. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt