grid-wise.us

Home Screen
Home Screen
Home Screen
Home Screen
Dashboard
Dashboard
Dashboard
Talk to agent

Inspiration

As AI and cloud computing continue to grow, so does their environmental impact. Many compute-heavy workloads such as model training, rendering, data pipelines, and batch jobs run without considering how clean or dirty the electricity grid is at that moment.

We were inspired by one simple idea:

$$ \text{Same Compute Job} = \text{Same Carbon Footprint} $$

The exact same workload can run at different times or in different regions and create very different emissions.

We wanted to build a platform that makes carbon-aware compute scheduling as easy as selecting a cloud region.

What GridWise AI Does

GridWise AI helps users run flexible compute workloads at the cleanest possible time and location.

Users enter:

Region / Grid Zone
Hardware Type
Job Duration
Earliest Start Time
Deadline

GridWise then:

✓ Finds the cleanest run window
✓ Compares against running immediately
✓ Calculates CO₂ savings
✓ Explains why the schedule is cleaner
✓ Suggests better global regions
✓ Provides AI chat assistance

How We Built It

We built GridWise AI using three major systems:

1. Backend (FastAPI)

Our backend connects to Electricity Maps live grid data and runs the optimization engine.

For a workload lasting n hours:

$$ \text{Best Window} = \underset{i}{\arg\min} \sum_{i=1}^{n} \text{CarbonIntensity}_i $$

This means we evaluate every valid time window before the deadline and choose the one with the lowest total emissions.

2. AI Agents Layer

We built a separate AI service powered by Gemma. Users can ask:

"Why did it choose this time?"
"What if we start one hour earlier?"
"Explain this simply"
"Compare baseline vs optimized"

The AI uses the real optimization output as context.

3. Frontend

We built an interactive dashboard using JavaScript + HTML/CSS + Chart.js.

The frontend displays:

Carbon timeline charts
CO₂ savings metrics
Region recommendations
Real-world impact figures
AI assistant chat

Machine Learning Model

We created Region Intelligence, an ML model that predicts the cleanest region for a workload.

We trained an XGBoost regressor on 106,000+ real-world samples across 25 global grid zones.

The model predicts: Average Carbon Intensity Over Job Window

This helps answer: Not only when should I run but where should I run?

Challenges We Faced

1. Live API Issues

We initially used the wrong API version, which returned missing data. We had to redesign our data ingestion pipeline.

2. CORS Deployment Problems

Our frontend and backend were deployed on different domains. Browsers enforce CORS (Cross-Origin Resource Sharing) — without proper configuration, the frontend could not call the backend API. We fixed this by explicitly allowing our frontend domain in backend settings.

3. ML Deployment

Training locally was easy. Production deployment required solving:

Python environment mismatches
XGBoost dependency issues
Model file loading
Memory limits

4. Chat Routing

We had to teach the system to distinguish between a question about a previous run vs. a request for a new optimization.

What We Learned

More than coding — Building production systems requires: APIs + ML + Deployment + Networking + UX
Data Matters More Than Fancy Models — Good features + good live data often outperform complex models.
Sustainability Must Be Easy — If reducing emissions is difficult, users ignore it. If it is one click, users adopt it.