# Luh Labs

- **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:** [Caleb Luh](https://cerebralvalley.ai/u/schmaleb)
- **GitHub:** https://github.com/Luh-Labs-LLC/muse-finetune-agents
- **Demo video:** https://youtu.be/mSki1-xyYpY
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4468

The common assumption about getting better answers out of an AI is that you need a better prompt. Most people have lived this. You ask for a sum of some cells, you get three paragraphs about spreadsheet best practices, and you type "no, just the formula" for the fifth time that week. The issue with this approach is that nothing is retained. Every conversation starts from zero, and the correction you made yesterday is gone today.

Trained Agents takes a different view. Instead of telling the model what you want every time, you teach it once. At agents.luhlabs.com you chat with a small model, and when it keeps getting something wrong, you type one sentence about what it should do. From there, two Muse Spark agents do the work that normally takes a machine learning engineer a weekend. The first, which I call the Job Architect, writes the training conversations, picks the base model, the LoRA rank and the number of epochs, chooses the rules the goal has to follow, and writes test prompts that it keeps out of training. Code checks every chat against those rules before any GPU time is spent. Then a LoRA adapter trains on Qwen3-0.6B on an RTX 4060 in under a minute.

To understand why the second agent matters, imagine a teacher grading essays with the names covered, reading each pair twice in a different order, and being allowed to say that neither essay answered the question. That is the Quality Judge. Both models answer prompts they have never seen, the judge picks the better reply twice with the order swapped, and only a verdict both readings agree on counts. When neither reply does the goal, it says so, and that counts against the fine-tune. Muse Wonder reads the same pairs as a second opinion. If the result is short of the bar, the agents write new data aimed at what was marked down, retrain, and judge again, up to four rounds. When they are done, the chat switches to the model they made.

Every score is public. Support tickets to one line of JSON: preferred over the base model told the goal on 19 of 20 prompts, no losses. Excel formulas from plain English: 11 of 19, with eight ties and no losses. A model taught from the chat to answer only in woofs and barks, in 40 seconds of training: 14 of 20. And one honest failure, writing like Mark Twain: 5 of 18, where the judge reported that neither reply did the goal on eight prompts. I left that run on the site on purpose. A pipeline that can only report wins is not measuring anything.

The problem it solves is simple. People correct their AI the same way every day. A small model that has learned your format beats a large one you keep re-prompting, it runs on hardware you already own, and with Muse Spark doing the design and the judging, teaching it takes one sentence.

*** FOR JUDGES: IF NEEDED, USE ACCESS CODE ABC123 TO TRAIN A MODEL ***

## More from Meta Global AI Developer Hackathon

- [Mykro AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4465)
- [sanjeev](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4466)
- [Itemized](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4467)
- [Troupe](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4469)
- [Metis](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4470)
- [Concourse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4471)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4468. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt