Aether

Aether predicts missing Scope 1–2 emissions with transparent ML and an AI assistant that explains every step at any expertise level, making ESG data clear, actionable, and trustworthy.

Comment

Aether: End-to-End Emissions Intelligence

Inspiration

Estimating emissions is hard even for experts and nearly impossible for smaller or non-reporting companies. Investors and analysts repeatedly told us:

“I want to price climate risk, but I can’t trust or even access the data.”

Aether was created to fill that gap: transforming messy, fragmented sustainability data into clear, decision-ready insight that anyone from a junior analyst to a chief risk officer can actually use.

What It Does

Aether estimates Scope 1 and Scope 2 emissions for non-reporting companies and makes the results explorable and explainable.

Under the Hood

  • Ingests:

    • Company fundamentals
    • Regional info
    • Sector revenue mix
    • Environmental activities
    • SDG commitments
    • (All from the Fitch dataset)

  • Predicts emissions using tuned ML models

  • Surfaces insights through:

    • A sleek, step-by-step B2B dashboard
    • An AI explainer assistant that:
    • Answers questions about data, pipeline, and model behavior
    • Adapts explanations to Beginner, Intermediate, and Advanced audiences

How We Built It

Data & Feature Engineering

  • Explored coverage and missingness across all Fitch tables
  • Engineered features from:
    • Region & country (proxy for grid mix + regulation)
    • NACE sector revenue distribution (sector intensity + diversification)
    • Environmental score + activity-based adjustments
    • SDG exposure indicators
  • Handled missing SDG data via KNN imputation
  • Used gradient boosting to impute environmental activity signals

Modeling

  • Built baseline models → tuned tree-based models for Scope 1 & 2
  • Logged:
    • Data splits
    • Hyperparameters
    • Metrics
    • Artifacts
  • Visualized baseline vs tuned performance with clear narrative lifts

Experience & AI Layer

Designed a narrative dashboard that guides users through:

  1. Data coverage & missingness
  2. Imputations (KNN + Gradient Boosting)
  3. Model training & validation
  4. Final emissions estimates & peer comparisons

AI Assistant grounded in:

  • Dataset schemas
  • Training logs
  • Diagnostics
  • Pipeline summaries

So it provides traceable, not generic, explanations.

Challenges We Ran Into

  • Accuracy vs interpretability
    Hard to tune models without making them “black boxes” for users/judges.

  • High-dimensional relational data
    NACE mixes, environmental activities, SDG commitments → meaningful features without noise.

  • Dashboard overload
    Early versions were “too much chart, not enough story.”
    We reorganized sections and added context text.

  • Expertise-aware explanations
    Making explanations useful to beginners and domain experts.

  • Deployment friction
    Packaging notebooks, models, and artifacts into a reproducible demo within hackathon time.

Accomplishments We’re Proud Of

  • Built a clean, end-to-end pipeline:
    Raw Fitch data → engineered features → trained models → emissions predictions

  • Achieved meaningful gains between baseline and tuned models and showcased the lift visually.

  • Designed a story-driven UI that guides users naturally from: data quality → feature engineering → modeling → emissions insights

  • Created an adaptive AI assistant that:

    • Explains predictions in plain language
    • Provides technical deep dives for data scientists
    • Surfaces limitations, contributions, and model behaviors

What We Learned

  • The importance of walking through the entire ML lifecycle, not just the final model.
  • Feature engineering is critical for sustainability data: sector, geography, and practices often matter as much as size.
  • Explainability + narrative matter just as much as RMSE for climate analytics.
  • AI assistants are far more useful when anchored to real pipeline artifacts, not generic LLM responses.
  • Clean narrative dashboards make complex ML pipelines accessible.

What’s Next for Aether

1. Multi-Agent Architecture

  • Ingestion Agent that pulls updated financials, sector mix, sustainability disclosures
  • Data Quality Agent that flags anomalies, missingness, outliers
  • Model Tuning Agent that retrains, benchmarks, selects models
  • Reporting Agent that drafts investor- or regulator-ready climate reports

2. Continuous Learning Loop

  • Collect human feedback on explanations
  • Retrain with new reported emissions + updated Fitch scores
  • Monitor drift and trigger automatic reevaluations

3. Deeper Scenario & Risk Analytics

  • What-if tools for:
    • Sector shifts
    • Environmental practice changes
    • Policy shocks
  • Incorporate grid carbon intensity trajectories for refined Scope 2 predictions

4. Integrations & Ecosystem

  • APIs into ESG platforms, portfolio systems, risk dashboards
  • Exportable audit trails (sources, model versions, explanations)

5. Richer AI Assistant Capabilities

  • Multi-step reasoning for complex questions:
    • “How does my portfolio’s Scope 2 exposure change if European grid emissions fall by 20%?”
  • Role-aware modes:
    Investor, Risk Manager, Sustainability Officer, Student
    layered on top of skill levels Beginner / Intermediate / Advanced

The Vision

Aether evolves from a hackathon project → into a continuously-learning, explainable emissions copilot powering the next generation of climate risk intelligence.

Built With

Share this project:

Updates