# Jeejak

- **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:** [Samsul Hadi](https://cerebralvalley.ai/u/samsul)
- **GitHub:** https://github.com/jeejak/proofmuse
- **Demo video:** https://youtu.be/OJSj9r7RKPY
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/50

ProofMuse checks an Indonesian online seller's packing photos while the parcel is still on the table, and says in Indonesian what to retake.

Sellers on Shopee and TikTok Shop lose disputes because their proof photo turns out useless: too dark, the shipping label cut off, the box never shown sealed, or the wrong label stuck on. They find out days later, when nothing can be re-shot. Jeejak's app JejakPaket records that evidence today: in its pilot since mid-September 2026 it has recorded 3,076 parcels and 1,545 packing photos in the cloud and synced 3,844 marketplace orders. ProofMuse is the new AI layer that checks that evidence at the bench instead of after the fact.

How it works: Muse Spark 1.3 labels every photo against a strict JSON schema. SAM 3.1 locates the shipping label on each photo. Where a label was found but not read, Muse Spark reads it again from a high-resolution SAM crop. Before the system may say "wrong label", it verifies: any tracking number that differs from the order is read a second time, and the mismatch stands only if both reads agree. Then deterministic Go rules, with no model, decide the verdict (Bukti kuat, Perlu dicek, Bukti lemah) and show fixed Indonesian retake instructions. The model is never told the expected tracking number.

Measured on a held-out test split of 62 parcels, labelled blind by a person: 100% precision and 100% recall on weak evidence, no false alarms on strong evidence, 60 of 62 verdicts in agreement, $0.014 and 4.6 s per parcel. Without the verify step, a one-digit misread became a false "wrong label"; verify removed it. The photos are synthetic (AI-generated staged packing with dummy waybills), so these are not real-photo numbers yet.

It is advisory: it flags evidence quality for a person to act on and never claims to prove what was in the box or decide a dispute.

After a complaint, the seller pastes the buyer's message: Muse Spark only classifies it into a claim type, and Go rules pick the photos that answer it and say what photos cannot show. One button then builds an evidence packet PDF from the exact photos checked: verdict, fixed reasons in Indonesian and English, the label boxes and each photo's SHA-256.

It already runs inside Jeejak's own product. During the hackathon we added ProofMuse as a backend for the Jeejak console's evidence check: photos synced from the JejakPaket app go to ProofMuse's API and the verdict shows in the console. It is off in production and runs on a demo deployment with synthetic data only, until the privacy policy names Meta (the console repo is private; the code is quoted in docs/console-integration.md).

Production pieces: async checks with idempotency keys over the order and photo hashes, per-client and daily rate limits, retries with Retry-After, per-call token and cost accounting, and a step trace on every result.

Built during the hackathon vs pre-existing: see HACKATHON.md in the repo. It builds on our earlier Gemini prototype ParcelProof (verdict rules and 279 eval photos ported and listed); the per-photo Meta pipeline, SAM grounding, reread, verify, retake loop, claim triage, evidence packet, server-to-server API and console integration, demo, 40 new eval records and the human labels are new. Added Nov 3: an open-weight comparison column, Muse Spark Open on Baseten, in the README results table.

## More from Meta Global AI Developer Hackathon

- [Plates for Hope](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/47)
- [Stephen Thuku](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/48)
- [Quantara](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/49)
- [Tendry](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/51)
- [Alpacha](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/52)
- [CoreTech Innovations](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/53)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/50. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt