Inspiration
Data centers are invisible megacities. A single large AI training cluster can draw 20–50 MW of continuous power—equivalent to a mid-sized city. Across Georgia, four hyperscale facilities (Google Douglas, Microsoft Fulton, Meta Henry, AWS Forsyth) now represent a measurable fraction of peak grid demand during summer afternoons.
When multiple data centers operate during peak hours simultaneously with consumer demand, regional utilities face a critical choice: either risk brownouts or activate legacy fossil-fuel peaker plants. These plants are catastrophically inefficient—they consume millions of gallons of water for cooling while operating at 30–40% utilization.
We realized this wasn't just a carbon problem. It's a grid reliability problem, and it's a water security problem. Digital workload scheduling decisions made inside a data center directly cascade into physical water stress on Georgia's local watersheds.
VIRIDIS exists to close this gap. We built a coordination layer that lets data center operators see exactly when their decisions impact grid stability—and what they can do about it in real time.
What it does
VIRIDIS is a Grid Stress Intelligence Platform that translates real-time power demand into operational grid risk, and offers operators an immediate mitigation path.
Core Features
Grid Stress Score (Real-Time) The platform calculates a dynamic stress metric (0–100%) representing current demand as a percentage of Southern Company's regional peak capacity (46,000 MW). This score maps directly to Georgia Power's actual tariff windows—specifically the critical summer peak (weekdays, 2:00 PM–7:00 PM), when peaker plant activation becomes likely.
Water Impact Modeling (Instant Consequence Translation) The system translates grid stress into ecological metrics, dynamically scaling water overhead based on operational state. At baseline load (grid stress ~28%), the platform calculates approximately 24,000 gallons of water required for cooling. During critical peak hours (grid stress 78%), this escalates to 150,000 gallons—visibly demonstrating to operators how their workload decisions directly cascade into physical water consumption by downstream utility peaker plants.
Workload Shift Mitigation (One-Click Closed-Loop Deferral) With a single master action toggle, operators initialize VIRIDIS closed-loop optimization mode. The system executes a deterministic state transition: compute-intensive workloads defer into Super Off-Peak windows (11:00 PM–5:00 AM), grid stress instantly drops from critical (78%) to safe (18%), and water overhead reduces by 82%. This closed-loop handshake is repeatable, auditable, and requires zero manual intervention.
Water Safety Index (WSI) Per Facility) For each monitored data center, VIRIDIS calculates a proprietary Water Safety Index (0–100) based on four EPA/USGS water quality metrics: pH, dissolved oxygen, temperature, and turbidity. The platform tracks 7/30/90-day trend predictions using scikit-learn models trained on real USGS watershed data, alerting operators when water stress indicators begin to degrade.
How we built it
We engineered a clean, decoupled full-stack architecture optimized for reliability under tight hackathon deadlines.
Backend: Deterministic Rules Engine + ML Forecasting Language/Framework: Python, FastAPI Grid Data Source: EIA (U.S. Energy Information Administration) hourly electricity API, filtered to Southern Company's RTO region Water Data Source: USGS EPA Water Quality Portal — real Georgia watershed measurements from 293 monitoring stations across Douglas, Fulton, Henry, and Forsyth counties Forecasting: scikit-learn time-series regression trained on 89+ months of historical water quality data; synthetic anomalies injected to validate early-warning sensitivity Logic Engine: Deterministic rules mapped to Georgia Power's actual peak-hour tariff windows, ensuring judges see reproducible, verifiable behavior API Contract: Clean REST endpoints expose grid stress scores, water safety indices, and mitigation state transitions as JSON, decoupled from frontend
Frontend: Interactive React Dashboard
Dynamic grid stress gauge (0–100%) WSI trends per facility (7/30/90-day forecast lines) 24-hour demand timeline slider Facility selector (clickable data center tiles) Operator/Community toggle
Challenges we ran into
Our initial architecture was too ambitious.
We attempted to build:
Live geospatial tracking (facility-level GPS coordinates) Full cryptographic data validation pipeline (SHA-256 per measurement) Real-time ML model retraining (every 6 hours) Multi-facility cyber-threat detection (suspicious power anomalies) Database persistence layer (PostgreSQL + ORM)
Accomplishments that we're proud of
✓ Bulletproof Full-Stack Architecture: Built a responsive, bug-free app handling complex state mutations flawlessly under deadline pressure.
✓ Domain Specificity: Instead of generic carbon tracking, we modeled logic around real utility infrastructure—Georgia Power's actual peak-demand windows, Southern Company's regional capacity benchmarks, EPA/USGS data standards.
✓ Accurate Data Integration: Pulled real watershed monitoring data (4,500+ pH readings, 4,500+ dissolved oxygen readings, 13,600+ turbidity readings) from USGS across four Georgia counties.
✓ Clean API Contract: Frontend and backend developers worked cleanly off a single JSON schema. Zero integration surprises at merge time.
✓ Transparent Methodology: Every number is traceable to a source (EIA, USGS, EPA, Georgia Power public capacity data). No black-box claims.
What we learned
In high-stakes engineering, scope control is everything. The most dangerous moment in a hackathon is when you're 60% of the way through and realize you're building five things at 60% quality instead of one thing at 100%. We learned to kill our darlings and commit to depth over breadth.
We also learned the physical infrastructure loops connecting digital decisions to ecological consequences. Cloud computing isn't abstract—it's millions of gallons of water, specific counties, real grid operators making real-time decisions. When you map the stack, you realize that API calls have watersheds.
Finally, we learned that accuracy is non-negotiable. Early in the sprint, we made claims about AI data center water consumption that didn't hold up under scrutiny. We corrected course, rewrote the pitch, and grounded every claim in published data. Judges can smell handwaving. Specificity wins trust.
What's next for VIRIDIS
Phase 1: Multi-Facility Orchestration Expand the dashboard to manage interconnected clusters across multiple regional nodes simultaneously. Imagine a single pane of glass showing Google Douglas + Meta Henry + Microsoft Fulton + AWS Forsyth all on one grid-balancing scorecard, with cross-facility workload optimization recommendations.
Phase 2: Predictive Grid Modeling Integrate weather forecasting (NOAA) and consumer demand prediction (Fed's economic indicators) to forecast grid stress 7–14 days in advance. Help operators plan major batch jobs with 2-week lead time, instead of hour-to-hour reaction.
Phase 3: Autonomous Grid-Shedding Oracles (Ongoing) Transition from operator-initiated toggles to fully autonomous, smart-contract-driven grid balancing. By exposing the telemetry API to edge nodes, regional facilities can execute split-second automated load-shedding commands when utility frequencies drop below critical safety thresholds, ensuring absolute resilience without a single millisecond of human latency.
Log in or sign up for Devpost to join the conversation.