# Argus Labs

- **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:** [Aymen Awel](https://cerebralvalley.ai/u/Akhi-Ay)
- **GitHub:** https://github.com/aymenawel/Argus
- **Demo video:** https://youtu.be/Lqan_2ra-RI
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4892

Project Description (changed: Meta models line)

Argus is an AI lab partner that turns any webcam or USB microscope into a scientific instrument. You set up your experiment, write down what you expect to happen, and point a camera at it. Argus watches, measures what changes in real units, judges your hypothesis and suggests the next run: Plan → Watch → Measure → Judge → Next run.

The problem: experiments are slow, and the moments that matter often happen overnight or between checks. Researchers end up measuring once at the end, judging by eye (“looks browner”), and writing notes from memory. AI that monitors and reasons about experiments exists today in multimillion-dollar self-driving labs and enterprise systems. Students, teachers and small research teams don’t have access to it. Argus brings it to anyone with a camera.

How it works:

Watch: a local watcher reads every frame and only wakes the AI when something actually changes.
Measure: SAM 3.1 outlines what’s in each container, and code turns those outlines into real measurements like length, area, colour change and counts, each with its margin of error from a printed scale card. The AI never makes up a number.
Notes: Muse Spark writes plain-language notes about what happened, and every number in a note has to match a stored measurement. A claim checker rejects any note that invents a number or claims more than the data shows.
Judge: the researcher’s plan is locked before the run starts. Code decides which verdicts the data allows (supported, not supported, inconclusive or too early), Muse Wonder explains the verdict and points out other possible causes like lighting or temperature, and a second review checks it.
Next run: Argus uses this run’s results to suggest what to test next and explains why.

Everything is traceable: the raw footage, the measurements, the AI’s notes and the researcher’s own reviews are kept separately, so nothing gets overwritten, and every AI call is logged with its cost.

In our own test, three glasses of cold, room-temperature and warm water with dye, Argus measured how much faster the colour spread in the warm glass, and every note it wrote passed the claim checker. It still refused to give a final verdict, because the plan was written after filming, and told us to lock the plan before filming next time. We also tested it on a published seed germination dataset, where it counted sprouted seeds against human labels.

Meta models: SAM 3.1 for seeing and outlining, Muse Spark for notes and turning the researcher’s own words into a plan, Muse Wonder for verdicts, reviews and next-run suggestions, and Meta’s open-weight Muse Spark Open, which also ran Argus’s notes through Baseten in our test runs.

Try it without installing anything at argus-replay.vercel.app: real recorded runs, replayed in the app.

## More from Meta Global AI Developer Hackathon

- [Musegix](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4889)
- [Janus](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4890)
- [One Cube Up](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4891)
- [Open Intelligence](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4893)
- [MIDUSAFE Sentinel](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4894)
- [Zān](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4895)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4892. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt