# Andaman Khunaprapakorn

- **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:** [Andaman Khunaprapakorn](https://cerebralvalley.ai/u/AndaK)
- **GitHub:** https://github.com/AMK-bot/StowLM
- **Demo video:** https://www.youtube.com/watch?v=fTb7Hn0OjTY
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3961

My family works in the shipping carrier industry: a razor-thin-margin business the global economy runs on, where every cost saving reaches producers and consumers worldwide. Central to it is the load plan, the schematic for how cargo goes onto a ship. A bad one means logistical inefficiency, extra crane fees, and fire or chemical hazards when dangerous goods sit too close.

Why do bad load plans matter? How much cost savings could there be?

Annually, across its impacts, faulty load plans cost companies an estimated 350 million to 6.6 billion USD. Our research places a moderate estimate at 2.4B across: Re-stow fees, Fuel Charges, Missed time slots, Delays, and Hazard damage from fires/chemicals/cargo instability. It is also estimated that these faulty plans also account for anywhere from 0.03 - 5.6 deaths per year

My solution: the StowLM pipeline.

StowLM is a pipeline, packaged as an app, that outputs a verified load plan and is built to cut down the time-consuming, error-prone work of developing one. Three Meta models do the heavy lifting. Muse Spark reads the paperwork (a photographed tally sheet, a booking, a shipping-instructions sheet), flags inconsistencies and, where something is ambiguous, asks a human supervisor. Muse Voice Transcribe turns a planner's radio call into text. StowLM Reader, a Muse Glimmer 30B model we fine-tuned locally with LoRA, transforms the planner's plain words, typed or spoken, into rules compatible with our plan-solving software.

For comparison, 47% of the rules outputted by an untuned glimmer 30B were accepted by our solver software. After LoRA tuning, this jumped to 92%.

To generate a load plan, we use Frontier Engine 0.13.0, a solver software we developed with coding agents, which searches for the most move-efficient plan that abides by every rule (standard shipping rules + added rules). Then a full checker is run on the result, from dangerous goods to stability and hull strength. This engine is intentionally left as an isolated module, so large scale companies could port their own plan-solver into the pipeline to allow for ease of adoption into the industry.

On test voyages, 24 of 24 plans passed the full checker. A simple two-step planner (outdated but once used piece of software), passed none.

The pipeline and the fine-tuned adapter are open source. Research demonstration on synthetic ships and cargo at a simulated port: up to 195 boxes in a half-ship section of 500–800 TEU feeders, up to 70 % fill; the engine isn't competitive at large scale yet. The base model's usage policy prohibits use related to operating transportation technologies where harm is possible, so this must not plan or operate real vessels; an operational version would need terms that permit it.

A field expert (Head of Operations at Evergreen Thailand) was consulted, he endorsed the project’s potential for real deployment.

## More from Meta Global AI Developer Hackathon

- [Second Opinion](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3958)
- [Tonelab Muse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3959)
- [CloveaXD](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3960)
- [GeoSAR](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3962)
- [Lean](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3963)
- [Spaces](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3964)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3961. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt