# Grudge

- **Event:** [The Harness Engineering & Model Wrangling Hackathon](https://cerebralvalley.ai/e/mongodb-nyc-hackathon)
- **When:** Sat, Sep 26 at 9:00 AM – 10:00 PM (EDT)
- **Where:** New York, NY
- **Team:** [Shikha Soneji](https://cerebralvalley.ai/u/shikhasoneji), [Sameer Neve](https://cerebralvalley.ai/u/sam280691)
- **GitHub:** https://github.com/shikhasoneji8/infra_design_permits
- **Demo video:** https://youtu.be/gp7qtiw8Nts
- **Gallery:** https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/9

1. The problem

Data center permitting in the US takes one to three years, and much of that time is rework. A site plan goes to the county, comes back with rejections (generators too loud at a home, cooling in a wetland buffer, NOx over the Title V line), and the design team starts over. The rejection letter is filed and forgotten. The next site, or the next round on the same site, repeats the same mistakes. Towns are responding by banning data centers outright; in New Jersey alone, six municipalities did this in 2026.

2. Existing solutions

Today this is consultants with CAD and spreadsheets, one site at a time, with no institutional memory beyond the people in the room. AI tools in the space either summarize regulations (RAG over permit documents) or generate plans without checking them, so an LLM grades its own homework and nobody trusts the score.

3. Our solution

Grudge is a long-horizon agent harness where the math is code and the judgment is AI. A designer agent places a 40 MW data center on a real parcel pulled live from state GIS (13 parcels today across New Jersey and Travis County, Texas, each under its own rulebook). A reviewer applies real state law with plain Python: noise propagation to every neighboring home, wetland transition areas, air permit caps, NOx potential to emit, water withdrawal, stormwater, setbacks. A penalty score computed by code says pass or fail. Every rejection becomes a lesson in MongoDB Atlas, recalled by Vector Search before the next design. The loop runs as a LangGraph checkpointed by Atlas: kill the process mid-run and it resumes from the same round. Rejection letters are indexed with Atlas Search as precedent; parcels, homes and wetlands are GeoJSON with geospatial queries. Harness rules, not model size, made it converge: never regress from the best plan, let code repair geometry the model gets wrong, break stalls with a deterministic policy.

4. Impact

On a parcel the harness had never seen, memory off took three review rounds; memory on opened at a third of the penalty and passed in two. Every site it designs makes the next one faster, and the output is a plan a permit officer can check line by line: air-cooled with low-noise fans behind a screen wall, Tier 4 silenced gensets behind a berm, all 40 MW retained. Fewer resubmissions, quieter neighborhoods, and a memory that outlives the team. The same harness applies to any process with a strict reviewer.

## More from The Harness Engineering & Model Wrangling Hackathon

- [Hindsight](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/6)
- [Agent Evolve](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/7)
- [SwipeHome](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/8)
- [Byte Club](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/10)
- [Da Vinci: Recursive Improvement CAD Harness](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/11)
- [Jev classifier + Citation Harness](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/12)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/9. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt