Phantom Busters

Dashboard
Decision Logs + Gemini and Elevenlabs powered AI Assist
Documentation
Dashboard (dark mode)
Data Pipeline Overview
Decision Logs Filter Demonstration
Dataset Overview (taken from GWA-T-12 BitBrains)

Track Chosen

We chose the “Smart Grid” track with an aim of making energy distribution more resilient and secure.

Problem Statement

Phantom loads occur when devices continue to draw power while in standby mode, silently wasting energy and increasing costs. As they are difficult to detect, phantom loads are often overlooked inefficiencies in modern systems. As Georgia Tech itself has noted, optimizing data center efficiency is now essential in the face of AI’s ballooning energy consumption.

Georgia Tech. (2025, September 3). AI’s ballooning energy consumption puts spotlight on data center efficiency. Georgia Tech News Center. https://www.gatech.edu/news/2025/09/03/ais-ballooning-energy-consumption-puts-spotlight-data-center-efficiency

Ideation & Development Process

We aimed to address energy issues that are relevant to a range of users from individuals to large data centers. Through initial research, we learned about phantom loads which silently drain power from grids, despite not needing the power. Through discussing our ideas with a keynote speaker and continuing our research, we learned that phantom loads are relevant in data centers where large requests are made without the intent of actually pursuing them. This results in high allocated power for the predicted future, which drives up costs and leads to miscalculations with power allocations. Through identifying phantom loads, it may be possible to mitigate these cost-concerns and poor allocation processes. We decided to create an AI agent that can track live energy data for inconsistencies in the base-level power used to identify phantom loads. Energy can then be diverted from the phantom server to another server which could work more efficiently with added power. In addition, our AI agent scans for energy allocation anomalies to determine potential physical breaches within the data center’s servers and alerts the user of these possibilities.

Solution Proposed and Intended Impact

Our solution was to create a website that would not only detect the occurrence of phantom loads in a datacenter, but also mitigate this issue through the use of an AI agent. Our website would also take the opportunity to search for any potential cybersecurity breaches as it detects anomalies in the datacenter’s energy usage. With our solution, we hope to increase the efficiency of modern datacenters and decrease their overall energy use as well as the cost.

Data and Sources

GWA-T-12 (BitBrains) is a publicly available workload trace dataset provided by the Grid Workloads Archive (GWA), a repository maintained by the AtLarge Research Group that offers anonymized performance traces from real grid and cloud computing environments for use in research and evaluation.

The dataset is split into two major trace collections, one of which we pulled data from to train our agent and machine learning algorithms. The Rnd trace dataset contains 500 VMs connected to a mixture of fast SAN and slower NAS (Network Attached Storage) devices, including a larger proportion of management or auxiliary machines.

“GWA-T-12 BitBrains,” Grid Workloads Archive, AtLarge Research Group, accessed [date]. Available: https://atlarge-research.com/gwa-t-12/

How It Works: Technical Overview

Data & Modeling Layer

This is where we process real workload data from actual data centers and train our detection model. We start with real VM traces from the BitBrains dataset, then create synthetic data center topologies (racks, power distribution units) to simulate realistic scenarios. A machine learning model (Random Forest) learns to distinguish between two types of problems: genuine security threats ("cyber" anomalies) and wasted energy ("phantom" loads). The model's features (CPU usage, memory consumption) are normalized and stored so they can be reused consistently.

Backend & API Layer

The application is built with Node.js and Next.js. This layer ingests workload data from CSV files, calculates energy metrics at the rack level, and accesses key information through APIs. The raw data is transformed into time-series metrics that power both the dashboard and the AI agent's reasoning.

AI Reasoning & Voice Layer

Google's Gemini API analyzes the system metrics and explains what's happening in English. We use Gemini's tools feature to keep its reasoning structured and aligned with what's actually happening in the system. ElevenLabs text-to-speech gives the agent a voice: a "ghost" character in the interface that narrates energy saving decisions as they happen.

The Frontend

The dashboard is built with React and Next.js, styled with Tailwind CSS, and uses Recharts for visualizations. It shows energy consumption over time, breaks down phantom vs. cyber threats, displays the impact of power-saving actions, and streams live decisions from the AI agent.

Deployment

The platform runs on Vercel and is designed to run on the server side. The modular design means it can easily scale to analyze more datasets or connect to live data from real data centers.

Gemini: Used as AI Reasoning Engine

In this project, Google’s Gemini API is used as the core reasoning component of the AI agent. Instead of simply surfacing raw metrics, Gemini processes time series energy metrics (CPU, memory, utilization, etc.) created by the backend. This setup enables a hybrid model: metrics pass through classical ML (Random Forest) for classification, and then Gemini interprets and contextualizes those results in human-readable insights.

Technical highlights:

Metrics → features normalized → passed to ML classifier Classification outputs → Gemini contextualization

Natural language explanations tied back to real timestamped decisions.

Uses the API’s “tools” mechanism to ensure that its reasoning is aligned with actual system telemetry rather than hallucinating context.

Eleven Labs: Voice Synthesis Layer

To augment the UI and human-machine interaction, the system integrates ElevenLabs’ text-to-speech (TTS) API. Its role is to: Convert Gemini’s natural language explanations into spoken narration. Create a “ghost” persona voice that narrates energy diagnoses, alerts, and mitigation actions live as they happen. Give users audio feedback synchronized with what’s displayed on the dashboard.