# Zynplex

- **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:** [Juan Camilo Atencia Amin](https://cerebralvalley.ai/u/Jkamilo-ZPX)
- **GitHub:** https://github.com/jcamiloaa/gridxray
- **Demo video:** https://youtu.be/V92ha_PqARM?si=dcxZ8lqPh4MofOkq
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1047

GridXray — Physical AI that discovers the hidden wiring of electrical distribution networks by experimenting on them.

THE PROBLEM
Electrical utilities do not know exactly how their distribution networks are connected today. Records drift with every repair and every switching operation, switch positions are documented once and then changed in the field, and the network that is actually energized is invisible from the map. Wrong topology means wrong fault location, unsafe switching orders, planning studies built on fiction and customers restored late. Finding the truth is slow and expensive: crews walking feeders and engineers operating switches one at a time. So the real question is not only what the network looks like, but which test to run next.

HOW GRIDXRAY SOLVES IT
GridXray hands that question to a team of AI agents working on a faithful simulation of the network, before anyone touches the real grid.

1. Observe — The agents receive only what a utility really has: where each substation is, its smart-meter readings (noisy, sometimes missing, sometimes faulty), its equipment and the documented switch positions. Optionally, the utility's existing plan, given blind: the agents are never told whether it is current, old or full of errors.

2. Plan — A Planning agent checks that plan against evidence that costs nothing (how meters move together over a day, geography), decides how far to trust it, and sets how many experiments the investigation deserves.

3. Experiment — An Investigator agent forms hypotheses about which substation feeds which, and designs experiments to tell them apart: a load step here, a switch operation there. Every experiment respects operating limits; an unsafe one is refused and the agent is told why and how to retry safely.

4. Verify in the field — When the meters cannot settle a question, the agent sends a field crew. The crew photographs the site, the equipment is identified in the photo and the switch position is read, flagging records that are out of date.

5. Adapt — If the field changes during the investigation (a meter burns out, a substation loses supply), nobody tells the agents. They notice it in the readings, set aside evidence that no longer holds, and the Planning agent re-plans: extra experiments, a crew, or where to re-test.

6. Conclude — The result is the network as it really is, with every connection marked as measured, inferred or uncertain, so crews are sent only where the evidence could not decide.

WHY IT CAN BE TRUSTED
The true network is sealed before the agents start and the seal is verified at the end. The agents never see the wiring. A checker with no AI in it scores every connection against the truth.

RESULTS
Scored by the independent checker across repeated runs on networks of different size and difficulty:
• Small networks (15 substations): 82% of the real connections on average in a fixed benchmark, and up to 100% when the utility's plan is current.
• A published industry benchmark network, with a wrong initial plan and a live fault mid-run: 11 of 12 connections (92%), the run shown in the video.
• Larger networks (30 substations, two feeders, 20% unmetered): 63% on average with 15 experiments, about 7 minutes per network; up to 77% with a larger budget.
• Harder 60-substation networks with a very small budget: about a third recovered, but 36 of 39 mistakes were flagged as uncertain.
The key point is honesty: when GridXray is not sure, it says so, so crews go only to the connections the evidence could not settle instead of walking every site.


WHAT'S NEXT
Several investigator agents working in parallel, coordinated by the Planning agent, each on an independent part of the network, so investigations get faster as networks grow.

## More from Meta Global AI Developer Hackathon

- [Linevora](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1044)
- [Montivo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1045)
- [HABITA AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1046)
- [PatientVoice](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1048)
- [Jeronimo Vasquez Gomez](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1049)
- [Pauline Estira](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1050)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1047. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt