# Surfside Software

- **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:** [Stephen Frederick](https://cerebralvalley.ai/u/sfred)
- **GitHub:** https://github.com/sc-frederick/moke-meta-hackathon
- **Demo video:** https://youtu.be/wl8wwq_Tlrw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4710

MOKE is a desktop app that reviews engineering drawing sets before they go in for permit. It reads every sheet, records each issue w/ its source evidence, and leaves the call to the engineer.

I'm an AEC engineer in Tampa. A drawing set is full of facts that have to agree across sheets & disciplines. When they don't, the agency catches it, the set goes back for another round, and that costs the client time & money.

We tried an early version on a real job, basically these review prompts in Codex. After about 2 weeks of internal revisions, we submitted a 13k sqft custom home. It came back after a day w/ 3 comments, and the permit was approved 1 week after submittal. A job like that would typically have had 3+ rounds w/ the county. That's one job on a different model, so it doesn't prove what MOKE does, but it's why I wanted to build it.

One click reviews every sheet for completeness, code, coordination & prior comments across 11 disciplines. Sheets it didn't get to stay visible as gaps.

Each finding gives its location, evidence & correction. On a 13-sheet commercial set, it caught LS-1 totaling 1554 SF of tenant area while E-1 calculates its load from 1462 SF.

The engineer marks each finding valid, false alarm, not applicable or duplicate, w/ a reason, and follow-up reviews read those decisions. When a revised set comes in, findings carry forward. On that same set, the recheck updated 46 of 47 findings and found 11 new ones.

On sets of 4+ sheets, Muse Spark 1.3 runs as a coordinator plus one reviewer per sheet. Each reviewer decides which details to zoom into, looks up code sections & runs calculations, then the coordinator looks for conflicts between sheets. SAM 3.1 segments drawn devices & fixtures, and a finding that states a device count is rejected until SAM has run on that sheet. SAM also zooms into each finding's problem area.

MOKE also runs on Muse Spark Open, Meta's open-weight model, thru Baseten (meta-models/muse-spark-open-1026). On Oct 31, it finished that 13-sheet set in 25 minutes & ~$4 w/ 65 findings. It's less steady than Spark 1.3: a later run looped on some sheets until I cancelled it, so each sheet reviewer now has a request cap.

In single runs on Oct 30, Spark 1.3 took ~7 minutes & ~$2 at list price on a 5-sheet civil set, and 14.3 minutes & $11.47 on a 29-sheet brewery.

Pi Durable journals each review, and a ledger records every paid call before it's sent, so an unknown outcome isn't resent automatically. A review can run locally or in a Modal sandbox and sync back. Installers are on GitHub Releases.

The code index is Florida only. Ultimately the engineer still confirms the wording behind each citation. I haven't scored every finding against the drawings, so I don't have an accuracy number yet.

The MOKE name, review prompts, and original Florida code reference predate the hackathon. I built this application during the event using my project-kit desktop starter template. The UI adapts Forge, a licensed design pack.

## More from Meta Global AI Developer Hackathon

- [Slaver](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4707)
- [metro's-seasonal_web_design](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4708)
- [What to eat!](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4709)
- [Hive](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4711)
- [https://dev.meta.ai/?team_id=1954565191920593&project_id=1611618747641366](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4712)
- [OvaScope](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4713)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4710. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt