# Yigitcan Kizil

- **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:** [Yigitcan Kizil](https://cerebralvalley.ai/u/yigitcankzl)
- **GitHub:** https://github.com/yigitcankzl/plant-agent
- **Demo video:** https://youtu.be/x8i93pU-uSo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2130

PlantAgent started with a question: if an engineer asks an AI whether a plant can run 20% harder, who checks the answer?

It is built for the process engineer who has to answer "can we change this?" before a management-of-change review. Plant changes usually fail across disciplines: the process works, then the electrical supply, the relief case or the operating envelope does not.

An engineer gives PlantAgent a process drawing, operating data and a target. Muse reads the source material, builds a draft simulator model, finds the bottleneck and proposes a physical change. It does not grade its own work. The proposal is applied to the model and re-solved by engineering tools such as DWSIM, pandapower and WaterTAP. When a check fails, Muse gets the failed number and tries again. Documented equipment ratings cannot simply be edited to make a design pass, and any missing vendor or field information stays visibly open for a person.

The result is not a chatbot answer. It is a Rev B drawing with the change clouded on the sheet, the calculations behind each check, a what-if slider on the live model, and a draft change package that moves through a reviewer and a separate approver. Approval freezes the reviewed files into a hash-chained audit record.

I tested this in several ways during the event:

• In a recorded run built around BASF's public DEXPI example drawing and an illustrative process model, Muse Spark Open tried four designs. The server rejected two of them before a high-capacity tray proposal passed the modelled checks. The vendor tray rating remains open.
• A second model, Muse Spark 1.3, then tried to break that BASF design with seven bounded operating deviations. DWSIM re-solved each one; all seven held, the worst at 96 % of the flood limit.
• In a blind study based on Shell Norco's published revamp, Muse saw only the before-state data. It chose the same broad revamp type used by the engineers: high-capacity trays in the existing column. Against the withheld after-data, five conclusions matched, three were partial, one missed and one was not predicted.
• In the salt-plant case, the process model reached the 120% target, but the illustrative electrical model put a transformer at 133%. PlantAgent therefore kept the design blocked instead of presenting the process result as a complete answer.
• On a reduced reconstruction of the 2006 European grid incident, the hosted endpoint refused the task under its content policy. Muse Spark Open completed the same checked loop, and the proposed switching passed every simulated single-line contingency.

Beyond these, the workspace holds 16 documented facilities across process plants, power grids and water systems, with 13 blind studies scored against engineers' published conclusions. Not everything passes: 23 of 71 recorded engineering runs ended with every check passing, and most of those still carry an open vendor or human item. I kept those numbers in the repository rather than showing only the best runs.

PlantAgent is engineering decision support, not autonomous approval. Its job is to make an AI proposal testable and to show exactly what still needs an engineer.

Try it: a hosted replay runs without any key at https://plantagent.63-182-151-178.sslip.io, or git clone https://github.com/yigitcankzl/plant-agent && cd plant-agent && ./start.sh (a Dockerfile is included). Live Muse calls use your own Meta API key.

## More from Meta Global AI Developer Hackathon

- [CariBarang AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2127)
- [NeuroForge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2128)
- [WIzmux](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2129)
- [Realroom](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2131)
- [team cheetah](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2132)
- [Kolmo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2133)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2130. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt