# Marc Donovici

- **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:** [Marc Donovici](https://cerebralvalley.ai/u/Marc-Dvci)
- **GitHub:** https://github.com/Marc-Dvci/atlas-groundtruth
- **Demo video:** https://youtu.be/Xe0iSV5qkO4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/99

Atlas is a fleet of AI agents that audits public spending from orbit: it checks from satellite images whether the public works a government paid for were built. Give it a list of contracts, an area, a deadline and a budget, by voice or text. It checks every contract, largest and most suspicious first, and returns a tamper-proof case file for each: built or not, where, when, on what evidence, signed by an investigator.

I audit for a living. Audits test samples, because there is rarely time to check every line. In some domains one miss is too many: a flood wall paid for and never built can leave a village under water. Atlas checks every line, for audit offices, inspectors general, donors such as the United Nations, journalists and citizens. Roads, schools or forests only need a new list and map; the satellite archives cover the planet.

To test it against ground truth, I ran it on flood control in the Philippines, where auditors publish per-contract findings. Of 9,855 projects recorded complete (₱547.6 billion), a government review listed 421 possible ghosts, paid for but not found, and 337 of 416 were at a different location than declared. Site visits take months: by January 2026, 14 were confirmed.

Atlas audited all 1,096 contracts in Bulacan province (₱72.5 billion) in two hours, for $13 of AI costs. At that rate, the whole country would take under a day and $120. It ranked 84 leads (₱5.5 billion). These are not accusations: each tells an inspector where to go first.

Three checks that show it works:
- It agrees with the auditors. On 22 contracts officials had ruled on, never shown their findings, Atlas spotted a problem with 9 of the 10 never built and cleared none (the tenth had no image after completion). Where officials found a structure, it never called a ghost.
- A second satellite agrees. On 251 vegetated banks where the AI saw a new wall, Sentinel-2, read by plain code, recorded the change on 241 (96%).
- A second model agrees. Run entirely on Muse Spark Open, which an audit office can host itself, it gave the same verdict on 11 of 12 contracts.

Beyond the test, it flagged 52 more probable ghosts (₱3.3 billion), with 739,000 people on low ground around them. Three appear in audit reports Atlas never saw; for the other 49, I found no published audit.

The hardest part is location: 46 published pins sit over 1 km outside the barangay their contract names. So Muse Spark reads the contract and traces the river stretch with map tools, and Muse Spark Open re-checks each flagged one. SAM 3.1 and Muse Spark read the site's images year by year, and a reading counts only when Muse Spark Open agrees. Fixed rules, not an AI, give the verdict: a ghost needs two independent image sources showing nothing. Muse Wonder writes the case file, citing sealed evidence in every paragraph.

Nothing is simulated. Live: 
https://rgt48790--atlas-live-web.us-east.modal.run audits any contract you pick, from scratch.

## More from Meta Global AI Developer Hackathon

- [sidonsoft](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/96)
- [PlainMed](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/97)
- [ClearDesk AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/98)
- [iussum-agent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/100)
- [BossoVoice](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/101)
- [Zero White Dreams](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/102)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/99. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt