# DockTwin

- **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:** [Day Nguyen](https://cerebralvalley.ai/u/dayn)
- **GitHub:** https://github.com/nduy1234/DockTwin
- **Demo video:** https://youtu.be/ZWaomFt8mq4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3357

DockTwin: 

Shipping delays cost an estimated $68B in 2025 (Otonomi, 2025 Supply Chain Disruption Market Report). At terminals like LA/Long Beach, a delay often starts small. The paperwork says a box is empty, but the scale says 10 tons. A reefer warms up while the carrier says it’s fine. A customs hold lands just as a gate lane goes down. Planners juggle these conflicts by hand, and the usual answer is blunt: hold everything or close the gate.

DockTwin turns a plain-English description, a spoken note, or a yard photo into a programmable 3D digital twin of a terminal. An AI agent then stress-tests that twin against conflicting real-world signals. Planners see which boxes to slow, reroute, inspect, or release before a disruption becomes a delay, and exactly why.

How we use Meta’s models:

Muse Spark turns descriptions into a structured twin spec using JSON-schema output with automatic repair. It also invents adversarial scenarios, writes test oracles (“a reefer above its setpoint must be inspected”), proposes rule fixes as diffs, and explains decisions. With vision, it reads real container codes inside SAM’s masks.
SAM 3.1 segments yard and gate photos into containers, cranes, and lanes that seed the twin.
Muse Voice Transcribe takes spoken input, biased toward terminal vocabulary like box codes and lanes.
Muse Spark Open on Baseten is an optional alternate scenario provider.
The project was developed with a coding agent running on Muse Wonder.

Engineering:
Our design rule is “models plan, the engine decides.” Rules are written in a small DSL that is interpreted, never executed as code. A seeded, deterministic simulation runs every scenario, so the same spec and seed always give the same result. A fix is marked VERIFIED only when re-simulation shows the bug is gone and nothing new broke. Safety floors, such as customs holds and reefer temperature limits, don’t depend on model output and can’t be switched off by a rule. Each decision picks the gentlest action on a ladder from watch to hold to hand off to a human. It also scores source trust (scale over paperwork), confidence, and cost of delay, and needs human approval when it matters. Every step lands in a hash-chained audit log, and every model call is logged. The project has 466 tests, time-budgeted model calls, and a production smoke test of all six routes.

Dual-use:
On an ordinary Tuesday, the same maker and test agent help game studios and robotics teams debug interactive 3D objects. For example, it catches an automatic door that chatters when someone stands still in its safety beam.

Impact:
Planners get fewer blunt shutdowns, and truckers spend less time idling at dead lanes. Cargo arrives unspoiled and on schedule, and port neighborhoods breathe less diesel. DockTwin fits Push the Frontier: a better way to understand and control a physical system.

## More from Meta Global AI Developer Hackathon

- [Steel AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3354)
- [Human](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3355)
- [Hyper](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3356)
- [Aga](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3358)
- [MuseOrg](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3359)
- [Cutting Edge: A Ninja Fighting Game](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3360)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3357. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt