# Glassline

- **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:** [William Voorsluys](https://cerebralvalley.ai/u/williamvoor)
- **GitHub:** https://github.com/williamvoor/glassline
- **Demo video:** https://youtu.be/Vi0hn2Rwim0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4606

Glassline turns plain-English questions about Postgres into interactive answers you can trust and keep exploring.

 Problem: LLM data answers fail three ways. Models retype query numbers into prose and get them wrong. Answers are black boxes with no SQL to check. Every follow-up click starts a new slow chat turn. Onboarding a new database is equally painful — undocumented schemas and
 unreviewed AI descriptions.

 We split the work. Offline, Muse Wonder prepares the dataset once: it inspects the read-only schema, drafts notes, glossary, comments and Boards, with every factual claim backed by SQL the tool re-runs. A person reviews and exports before Glassline ever sees it. Nothing
 writes to the database.

 Live, Muse Spark answers: it writes read-only SQL under guardrails (single SELECT, allow-list schemas, timeout, 5k cap), stores Results as r1, r2, then writes layout referencing Result IDs — never typing values. The server expands Macros (BarChart, DataTable, Metric, Cite)
 from stored rows and shows the SQL under every visual. Cached prompts keep typical answers to 2 calls, ~6s median.

 Exploration then needs no model: Boards, Controls, Drill-downs and Entity Boards re-run prepared queries in ~0.01s with tabs and an investigation log.

 Parallel track, not in demo: Glimmer tuning for compact A2UI. LoRA SFT of Muse Glimmer 30B on Fireworks serverless (rank 16, 325 steps → s225) cut outputs 792→332 tokens. Muse Spark Open via Baseten was tested as teacher/judge for distill-and-filter best-of-8. Quality gap
 was inconclusive (0.505 vs 0.523 F1), driving new filtered data. Built with A2UI. Demo runs Spark-only on AusTender and VicProp.

## More from Meta Global AI Developer Hackathon

- [Zeno](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4603)
- [Emergency Medicine AI Research Initiative](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4604)
- [Anand Kumar (Solo)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4605)
- [Teleion Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4607)
- [RuleRepro](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4608)
- [ConformScope](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4609)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4606. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt