# AnaliZac

- **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:** [Raul Alejandro Aquino Ibarra](https://cerebralvalley.ai/u/Raul99)
- **GitHub:** https://github.com/Raul99Alejandro/coresight-hackathon
- **Demo video:** https://www.youtube.com/watch?v=BceyZbq7P9c
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1093

CoreSight takes a photo of a drill core box and returns a geotechnical and geological log per interval. Meta's models see the core, deterministic code does the math, and a geologist reviews the values and signs off each box. When a photo is not a core box, it refuses instead of measuring.

Live demo, no sign-up, 60-second tour: https://coresight.analizac.com

The problem: after every drill hole, geologists log core boxes by hand. A geologist logs at most about 5 m of mineralised core per hour (Marjoribanks, an author's blog), so one tray of 5 m takes at least an hour. We build in Zacatecas, which produces 41.8% of Mexico's silver, and Mexico made 20.4% of the world's silver in 2025 (Camimex, Informe Anual 2026). Before CoreSight, I built AI agent systems, ERP and CRM for mining contractors.

How it works:
1. OpenCV straightens the tray and sets the scale.
2. A gate checks calibration, Muse Spark's look at the whole photo and SAM 3.1's confidence. A screenshot or core with no tray is refused with its reasons.
3. SAM 3.1 cuts each piece of core. Where SAM's text prompt finds nothing, Spark's boxes recover the row.
4. Deterministic code turns pieces into RQD, recovery and fracture frequency per run. No model touches the number.
5. Muse Spark describes lithology, alteration and mineralization, validated with Pydantic. The same contract also runs on Muse Spark Open on Baseten and on Muse Glimmer on our own laptop.
6. A geologist corrects breaks and pieces by tapping the photo, by typing or by voice. The original stays next to the correction.
7. Approved boxes join by depth into one drill hole, with weak zones flagged by code and a 3D of the hole wrapped in the real box photos. The 3D trace is an example: public holes have no survey.
8. On approval, a hole agent reads the trends, Muse Wonder routes the box, and area agents draft for lab, safety and geotechnics, plant and finance. Every step is verified by code and then by Muse Spark as a judge, with at most three attempts before it goes to a person. A person approves every draft and nothing is sent.

Results: RQD mean absolute error 11.2 points, 95% CI 6.1 to 18.9, over 34 runs in 11 boxes, against a mine geologist's own log. Limits of agreement -40 to +26: large enough that a person must review the values before signing off the box. We do not claim it is within human variability. Under a minute and about US$0.05 per tray with the hosted models, before the agents; Spark's boxes and the agents are not priced, and the geologist's review time is not measured yet.

Limits: no mine or mine geologist has used it. The demo uses public photos. Lab and finance run when sulfides are visible, and no openly licensed full tray we found shows them, so the demo shows them as "Not needed"; their tests are in the repo. The learning loop showed no measured gain, so we do not claim one.

Built before the hackathon: five commits before kickoff, four on Oct 25 and one on Oct 26 at 11:45 AM ET, with the project scaffold, data contracts, the geometry formulas with tests, a survey of public photos and an API client not yet run. They stay in a separate public repo, Raul99Alejandro/coresight, pinned at commit 14bc545. The submitted repo holds only event work: install.py fetches that commit and applies our changes. Built during the hackathon: everything after commit 14bc545, including calibration, the gate, segmentation, descriptions, the review page, the drill hole view, the agents, the evaluation and the demo. The full list is in the README.

## More from Meta Global AI Developer Hackathon

- [First Pass](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1090)
- [EdAI Systems](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1091)
- [PackMatch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1092)
- [Koanic](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1094)
- [Musavir](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1095)
- [GT TechVerse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1096)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1093. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt