# ModelicAI

- **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:** [Anurag Dogra](https://cerebralvalley.ai/u/AnuragD2192)
- **GitHub:** https://github.com/anuragdogra2192/ModelicAI
- **Demo video:** https://youtu.be/HwctvjVfJj8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3864

ModelicAI Studio turns a plain-English description of a physical system into a verified, running simulation.

THE PROBLEM
Engineers model physical systems (machines, fluid circuits, motors, vehicles) in Modelica, the open standard for system simulation. It is powerful but hard: you wire components by hand from thousands of library classes, every class and parameter name must be exact or nothing compiles, solver errors like "initialization failed" take hours to debug, runaway simulations freeze your tools, and the learning curve is steep, so the know-how sits with a few experts. General AI tools make this worse: they invent component names and parameters that look right but never compile, and nobody verifies their output.

THE SOLUTION
ModelicAI Studio is a modelling workspace with verifiable AI agents built on Meta Muse:
1) Modi (Muse Spark 1.3) is the assistant: it understands the request, plans, explains every step and result, and calls the tools.
2) The designer (Muse Wonder 1.0, medium reasoning) writes the model from the real Modelica Standard Library 4.1.0.
3) A knowledge graph of the real library (4,606 classes, 23,668 parameters and connectors) checks every model before it runs, catching invented classes, wrong parameter names and mismatched connectors.
4) OpenModelica then compiles and test-runs each draft. If anything fails, the agent gets the exact cause (the variable and value that failed, not a generic error) and fixes it, for up to 5 attempts.
5) Nothing reaches the user unverified, and nothing is applied without their approval. Changes are shown as a reviewable proposal.
6) Results appear live on the diagram, with plots, run comparison and safe limits (progress, timeout, size cap, Stop).
7) Every agent step, Muse call and tool call is traced in Arize Phoenix.

From one sentence such as "a 1 kg mass on a spring and damper, start at 0.1 m, simulate 5 s", the user gets a compiled model, a simulation and an explanation of the result. They can then simply ask for changes like "add another spring in series".

WHY IT MATTERS
It removes the wiring work and the learning curve, and it makes AI output trustworthy for engineering, because every model is checked against the real library and the real solver before anyone sees it.

Open source (MIT), by Anurag Dogra.

## More from Meta Global AI Developer Hackathon

- [WarmStart](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3861)
- [Artrelo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3862)
- [mercymek](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3863)
- [Stock IA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3865)
- [TWIN](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3866)
- [UNTORN](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3867)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3864. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt