# ORCA by IssueLab

- **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:** [Dodzi Agbenorku](https://cerebralvalley.ai/u/dodzi)
- **GitHub:** https://github.com/IssueLab-Inc/ORCA
- **Demo video:** https://youtu.be/hoYnbPvhq0M
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3002

ORCA (Observe · Resolve · Check · Apply) is an AI design coordinator for buildings and machines. Most construction projects, and most machines, begin as 3D models made by separate teams. Each works on its own; brought together, parts end up in the same space. These clashes are found late, argued over for days and fixed by hand. Today's tools stop at a list of problems. ORCA proposes the fix as a real edit to the model, and proves it with geometry before anyone sees it.

- Observe: ORCA checks every model against every other with its own geometry engine. On Northbank, our demo office building, that is 464 clashes among 1,732 elements in about 4 seconds, each numbered and titled. Muse Spark Open, Meta's open-weights model, triages them all and finds the handful that need a fix (5 of 464 in our recorded run).
- Resolve: Muse Spark reads the clash from three rendered views and the model data (sizes, distances, openings it could use, the requirements) and proposes one to three edits to the real IFC elements: a move, an offset with bends, a trim, a reroute, a new opening, a resize, or a combination.
- Check: ORCA tries each idea in the model. Is the clash gone? Is the clearance met? Is nothing new hit, and does every joint hold? A failed idea goes back to Muse Spark with the reason. Only verified options reach the user. When Spark finds nothing, Muse Wonder takes a longer second try.
- Apply: the user compares before and after in 3D, beside plans, sections and elevations, and accepts. The server checks again, then edits the model. The edited IFC or STEP downloads with a PDF report and BCF issues for each team.

Around the loop:
- Show it built: Muse Image paints a checked fix as a picture of what will be built, Muse Wonder judges it against the model, and the model's own lines go over the picture.
- Workings: a live graph of every model call and every check.
- Insights: root causes and patterns across the models.
- Fixing several clashes together, and requirements that every fix must meet.
- Hand edits with undo.
- Filters by any property.
- Voice through Muse Voice Transcribe.
- An MCP server with pairing links, so your own agent, like Muse Code, runs whole workflows through the same check.

Built during the hackathon by IssueLab: React, TypeScript and three.js in the browser; Python, FastAPI and IfcOpenShell on the server; deployed on Modal; 934 server tests and 314 web tests. Northbank is a synthetic demo building IssueLab made before the event, used as data; all of ORCA's code was written during it.

## More from Meta Global AI Developer Hackathon

- [Uptrugime](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2999)
- [Univerlab.org](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3000)
- [ENFPDEV](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3001)
- [Etoundi Ondoua Ngando Emmanuel Tresor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3003)
- [Two Eyes](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3004)
- [Recognized](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3005)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3002. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt