Joulewise

Inspiration

India adds roughly one new electricity consumer every second — yet 685 million people globally still live without reliable power. The technology to change that exists. Solar panels are cheaper than ever. Government subsidies are available. Installers are ready. So why aren't more households and communities making the switch?

The answer isn't money. It's trust and information. When we talked to households, school administrators, and NGO workers considering solar adoption, we kept hearing the same fears: Will it actually work on my roof? Will I ever recover my investment? Which vendor can I trust? These aren't irrational fears — they're the result of a market flooded with salespeople and starved of honest, localized, data-driven guidance.

We built Joulewise because we believe AI shouldn't just serve the well-resourced. It should be the great equalizer — giving a household in rural Rajasthan the same quality of analysis that a large corporation gets from a team of energy consultants.

What it does

Joulewise is an AI-driven platform that takes three simple inputs from any user — their location, monthly electricity bill, and budget — and returns a complete, personalized renewable energy adoption plan in under 500 milliseconds.

A typical Joulewise output includes:

Best renewable configuration (Solar / Wind / Biomass / Hybrid / Wind+Storage) matched to the user's specific site conditions
Recommended system capacity and layout based on roof or land area
10-year ROI forecast with confidence bands (±6.2 percentage points)
Payback period and projected annual savings in rupees
Applicable government subsidies and private financing schemes
A go/no-go recommendation with an explainable confidence score

The platform is modular. Solwize handles solar-specific analysis including shading, panel degradation, and orientation optimization. Windwize evaluates wind feasibility based on local speed indices and terrain. Both modules share a common AI backbone, enabling users to compare energy sources side-by-side on a single dashboard.

Crucially, every recommendation is explainable. Joulewise doesn't just say "go solar" — it shows users exactly why, using plain-language breakdowns of the factors that drove the recommendation. This is central to our mission of fighting misinformation and building community trust in clean energy.

How we built it

The core of Joulewise is a two-model machine learning pipeline built in Python 3.10+ using scikit-learn:

Random Forest Classifier (150 estimators) Classifies the optimal renewable configuration from five possible output classes. Trained on ~2,000 synthetic samples mirroring Indian LMIC conditions.

# Classification: Best renewable type
clf = RandomForestClassifier(n_estimators=150, random_state=42)
# Accuracy: 0.89 | F1-macro: 0.87

Random Forest Regressor (200 estimators) Predicts 10-year ROI with confidence bands to communicate uncertainty to non-technical users.

# Regression: 10-year ROI prediction
reg = RandomForestRegressor(n_estimators=200, random_state=42)
# R²: 0.92 | MAE: 6.2 percentage points | RMSE: 8.1

The pipeline ingests 14 input features across four categories:

Category	Features
Geographic	Latitude, Longitude
Site	Roof area (m²), Land area (m²)
Demand	Monthly bill (₹), Peak load (kW)
Financial	Budget (₹), Max EMI, Desired payback years
Grid	Outage hours/day, Tariff (₹/kWh)
Derived	Solar potential, Wind speed, Biomass index

Training data covers realistic Indian conditions: bills from ₹1,500–₹30,000/month, budgets from ₹100k–₹1.5M, and grid outages from 0–12 hrs/day. The dataset is fully synthetic with no PII, geographically constrained to Indian coordinates (8–37°N, 70–90°E).

The web interface is built on Streamlit, with an interactive scenario copilot that lets users adjust parameters in real time. Models are serialized with Joblib for fast cold-start inference (<500ms per prediction).

Challenges we ran into

Data scarcity was our biggest obstacle. Real household-level energy data in India is fragmented, inconsistently labelled, and often inaccessible. We made the deliberate decision to train on carefully constructed synthetic data rather than risk training on biased or unrepresentative real data — but this means our validation remains theoretical until we run field tests.

Explainability was harder than expected. Random Forests are interpretable relative to deep learning, but translating feature importances into plain-language explanations that a non-technical rural user can act on required significant iteration on the UI layer.

Scope discipline was a constant challenge. Joulewise could expand in many directions — community microgrids, carbon tracking, policy recommendations. Keeping the prototype focused on the single-user, single-site case was something we had to actively enforce.

Accomplishments that we're proud of

1. End-to-end working prototype — real inputs, real recommendations, <500ms inference
2. Strong model validation — 89% classification accuracy, R² of 0.92 for ROI regression
3. Accessible UX designed for non-technical users with plain-language explanations
4. Modular architecture (Solwize + Windwize) enabling domain-specific depth
5. Responsible AI by design — no user data stored, stateless predictions, confidence scores always surfaced

What we learned

We learned that responsible AI for underserved communities requires humility at every stage. It's not enough to build something technically accurate — you have to build something people can interrogate, challenge, and trust on their own terms.

The most important design decision we made wasn't a model architecture choice. It was deciding to always show confidence bands alongside predictions rather than presenting a single number as fact. Overconfident AI in high-stakes decisions causes real harm — especially for households making a ₹500,000 investment on limited income.

We also learned that the hardest problem in community-facing AI isn't the algorithm — it's the last mile: getting clean data in from a household with inconsistent billing records, a non-standard roof, and limited internet access.

What's next for Joulewise

Field validation Run Joulewise with 20–50 real households and community organizations in India to validate model outputs against actual post-installation data, and move from synthetic training data to real-world fine-tuning.

Mobile-first delivery Migrate from Streamlit to a FastAPI + React Native stack to reach users on low-end Android devices with intermittent connectivity — the most common computing context in the communities we serve.

Windwize expansion Wind and hybrid configurations offer better ROI in many regions. Windwize needs richer wind speed datasets and terrain-aware modeling before it's production-ready.

Bias auditing Build fairness evaluation frameworks to ensure Joulewise serves all communities equitably — not just those with the best data coverage.

Nature gives us the energy. Joulewise ensures we don't waste it.