# By my self

- **Event:** [The Harness Engineering & Model Wrangling #SFTechWeek Hackathon](https://cerebralvalley.ai/e/mongodb-sftw-hackathon)
- **When:** Sun, Oct 11 at 9:00 AM – 10:00 PM (PDT)
- **Where:** San Francisco, CA
- **Team:** [Azita Darvishi](https://cerebralvalley.ai/u/1122)
- **GitHub:** https://github.com/AziiDarvii/gridmind-ai
- **Demo video:** https://github.com/AziiDarvii/gridmind-ai/releases/download/v1.0/GridMind_AI_Demo.mp4
- **Gallery:** https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/54

GridMind AI is a persistent, self-improving and governed multi-agent harness on MongoDB Atlas that turns a failing power grid into an independently verified fix.

Problem: AI data centers add megawatts to distribution grids built for homes and shops. Engineers must find where limits break, why, what to change, and prove the change works.

What it does: 10 specialized agents (circuit modeling, power flow, violation detection, root cause, corrective action, data center, optimization, independent verification, recursive learning, engineering report) analyze the public CIGRE MV benchmark network with a 4 MW AI data center. An AC power-flow solver finds 17 violations at the evening peak (14 undervoltages down to 0.895 pu, 3 overloaded transformers). Agents diagnose each one with numbers, size 19 remedies by simulation and evaluate 179 plans. An independent verification agent reloads the model from MongoDB and re-solves it at peak and light load before anything is labeled fixed. Result: a $62,000 plan (capacitor bank, transformer tap change, shifting AI-training load, battery dispatch): 17 violations resolved, 0 remaining, 0 new; losses down 26%.

MongoDB Atlas is the harness backbone: circuit model, every simulation result, persisted run state (runs survive restarts and resume), audit logs, and a recursive-learning memory of verified lessons that later runs retrieve (our AI-surge scenario reused the first run's lessons and was verified again). An interactive single-line diagram shows original, proposed and verified states. The language model never calculates electrical values; it only explains solver evidence.

Built with: MongoDB Atlas, Python, FastAPI, pandapower, SciPy, React, TypeScript, React Flow, Claude API. Public data: CIGRE benchmark, SimBench load profiles, Open-Meteo weather.

## More from The Harness Engineering & Model Wrangling #SFTechWeek Hackathon

- [Morph](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/51)
- [Vivy](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/52)
- [SREvolve](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/53)
- [PaperTrail](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/55)
- [Long Horizon Quant Research](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/56)
- [LLM deployment agent](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/57)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/54. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt