GridPilot
GridPilot: Multi-Agent Smart Message Gate for DISCOMs
Built at AgentX Hackathon 2026
Team PRATIK'S TEAM: PRATIK PANDA, Sparsh Sharma
Built with Multi-Agent AI, Artificial Intelligence, Machine Learning, LightGBM, Energy Management, Smart Grid, DISCOM, Explainable AI, FastAPI, Human-in-the-Loop, Time Series Analysis, Decision Intelligence
About the project
GridPilot is an intelligent, multi-agent decision-making platform designed to help electricity distribution companies (DISCOMs) evaluate operational alerts before they trigger costly actions.
The platform uses four specialized AI agents:
Fixed-Limit Agent: Detects abnormal deviations from historical load baselines.
Weather-Adjusted Agent: Identifies load anomalies after accounting for weather conditions.
Peer Agent: Benchmarks buildings against comparable peer groups.
Trend Agent: Detects persistent multi-day changes in consumption patterns.
Their outputs are evaluated by a trained LightGBM multiclass decision gate, which classifies signals as beneficial, harmful, or negligible using counterfactual economic cost evaluation.
An intelligent routing layer converts these predictions into four operational outcomes: ACCEPT, IGNORE, DISCOUNT, and VERIFY. Uncertain decisions are escalated through a human-in-the-loop approval workflow, preserving operator control over consequential actions.
GridPilot uses the BDG2 Robin commercial-building dataset, containing 51 buildings and 17,544 hourly observations. It features a FastAPI backend and an interactive operations dashboard with live simulation, automatic demo playback, agent-level insights, explainable ML predictions, historical replay, and configurable operational settings.
Our goal is to move beyond isolated anomaly detection toward explainable, economically aware, multi-agent intelligence for smarter and safer grid operations.