# MkDev

- **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:** [Muthukumar Arunachalam](https://cerebralvalley.ai/u/MkDev)
- **GitHub:** https://github.com/mk-tdev/crisiscraft
- **Demo video:** https://youtu.be/lVq9Kn6cOXw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/485

CrisisCraft — Practise the crisis before it happens. Replay the decision you wish you’d made.

A volunteer coordinator faces a flooding town: a bridge is closing, a shelter cannot unload, and three teammates hold different pieces of the truth. You give an order, resolve conflicting reports, approve a plan and watch the vehicles move. Then rewind one decision and see two futures side by side.

The MVP has two playable 3D drills: a focused tutorial and a 50-minute regional scenario with four destinations, returning trucks, a four-crate rescue helicopter, tides, gusts and a false landslide report. Meta’s Muse Spark powers Priya (logistics), Malik (shelters) and June (field conditions). Each receives only their own briefing; facts reach the planner through the radio. The crew uses tools and peer review to propose a dispatch. Deterministic code controls inventory, capacity, routes, deadlines and the truck reserve.

The signature moment: the hill route supplies the school at minute 8; choosing the faster bridge route triggers a closure and diversion, arriving at minute 11—two minutes late. Replay synchronizes both timelines against the same scenario events and starting resources. An after-action review links its explanation to recorded engine events. An Agents panel exposes model turns, briefings, tools and timings; a Meta-powered Autopilot can take command, with manual takeover.

Evidence beyond a polished demo: in 30 recorded live crew evaluations of the tutorial, the general order met the school deadline in 10/10 full-radio runs and 0/10 runs with June silenced. A further 10 explicit-bridge runs were late despite the crew voicing the warning. These are measurements of one authored scenario, not a general benchmark or proof of training effectiveness.

Built during this hackathon: scenario logic, simulation, prompts and orchestration, procedural world, command interface, replay and debrief integration. Third-party models, libraries and credited assets are documented separately.

Play: https://crisiscraft.vercel.app
Evaluation: https://github.com/mk-tdev/crisiscraft/blob/main/docs/eval/crew-eval.md
Contributions: https://github.com/mk-tdev/crisiscraft/blob/main/docs/contributions.md

Designed for coordination rehearsal, not real emergency guidance.

## More from Meta Global AI Developer Hackathon

- [KISKY TEAM](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/482)
- [Amr Elhabbal - Neura](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/483)
- [Doubloon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/484)
- [GROWHAND](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/486)
- [Clarity Room](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/487)
- [Handoff Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/488)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/485. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt