Applied AI research
Multi-agent AI platform for grid planning and large loads
A research prototype in which four specialist agents run grid simulations and search a hybrid-retrieval corpus, with every answer scored on whether its figures trace to a tool result.
Cover illustration generated with AI
System architecture
Plan the study, then hand each task to the right specialist
The problem
A single large load request, a data center asking for several hundred megawatts, sets off a chain of questions that no one specialist can answer alone: where the request sits in the interconnection queue, which network upgrades it triggers, which regulatory process governs it, and what the hourly system looks like once the load is energized. The answers live in queue data, planning studies, tariffs, and production cost models, and assembling them takes weeks of expert time. The queue keeps growing while that work happens.
What I built
I built a multi-agent AI copilot to support power grid planning and large-load integration, with a particular focus on AI data center siting. An orchestrator translates planning questions into structured study plans, identifies missing assumptions, and delegates tasks to four specialized agents covering interconnection, power system analysis, simulation, and regulation.
The agents use engineering tools built on pandapower to perform power flow, N-1 contingency, short-circuit, and site-screening analyses. Through the Model Context Protocol (MCP), they receive structured simulation results, allowing them to adjust assumptions, rerun analyses, and compare alternative grid connection scenarios; finding the largest load a bus can host, for instance, is a search over whether the power flow still converges. The system also retrieves relevant information from a technical knowledge base using hybrid search and GraphRAG, drawing on 94 technical and regulatory documents.
To test whether the system could carry out published research rather than answer isolated questions, I used it to reproduce the site-screening method of Kim et al. (Nature Communications, 2026) on the RTS-GMLC test grid. Across all 73 buses, accepting curtailable service raised the number of qualifying sites by 22% at both 300 MW and 900 MW, consistent in direction and magnitude with the gains the paper reports on a 2,000-bus Texas system. The reproduction also exposed a measurement problem: the test grid already fails 37 of its 105 single-line outages with no data center connected, so an absolute pass rate would reject every site for reasons unrelated to the new load. The system therefore scores each site against the same network without the load, which is the comparison an interconnection study actually asks for.
Each analysis produces a traceable record of tool calls, assumptions, and results, making it possible to review recommendations, identify incomplete checks, and rerun studies with modified inputs.
The platform evolved from an earlier prototype, GRID-FORGE, which used a rule-based router to assign tasks. The current version replaces that approach with a planning orchestrator capable of coordinating multi-step analytical workflows.
What the demo shows
The video walks through a complete data center grid connection study, from the initial question to the final recommendation. A developer proposes a 2,000-rack AI data center campus in California, with an estimated facility load of 249.6 MW, and is open to curtailable electricity service. The orchestrator develops a study plan and coordinates specialist agents to screen 73 buses in the RTS-GMLC test network, followed by AC power flow and N-1 contingency analyses at three shortlisted sites.
A monitoring dashboard tracks the study plan, agent activities, assumptions, and missing inputs throughout the workflow. When the network lacks the source impedance data required for short-circuit analysis, the tool explicitly flags the limitation rather than generating an unsupported result.
How I validate the results
Reliable engineering analysis requires more than plausible AI-generated explanations. I developed a deterministic evaluation framework that uses recorded execution traces to assess whether the agents follow a sound and verifiable analytical process.
The framework checks five aspects of each workflow: whether engineering quantities are supported by simulation outputs, whether tasks are delegated appropriately, whether retrieved sources are cited, whether new loads are evaluated against baseline network conditions, and whether agents use simulation feedback to refine their analyses.
These checks help identify unsupported conclusions, incomplete tool executions, and inconsistent analytical workflows. I also use the evaluation results to compare model configurations and monitor execution costs, allowing less computationally intensive models to handle tasks where they maintain comparable performance.
The evaluation measures the traceability and consistency of the analytical process, rather than guaranteeing the correctness of every engineering conclusion. Final results still require domain-specific validation and engineering judgment.
Limitations
The platform is designed to support, rather than replace, professional power system planning and engineering judgment. Its current simulations use publicly available test networks, including RTS-GMLC, rather than utility-specific grid models. Where essential inputs are unavailable, such as source impedance data for short-circuit calculations, the system flags the limitation instead of producing unsupported estimates.
The quality of document-based responses also depends on the coverage and currency of the underlying technical knowledge base. Consequently, the platform is best suited for preliminary grid connection screening and scenario exploration, with detailed interconnection studies and final engineering decisions requiring further validation.