Kepler Exoplanet Oracle

Interactive sci-fi dashboard connecting to the FastAPI backend. Users input Kepler telemetry for instant AI-driven classification.
Confusion matrix demonstrating our Random Forest model's 95% accuracy across Confirmed, Candidate, and False Positive classifications.
SHAP interaction plot providing Explainable AI transparency. It maps exactly how specific Kepler telemetry influences model predictions.

💡 Inspiration

NASA's Kepler space telescope gathered vast amounts of planetary light-curve telemetry, producing far more candidate objects than astronomers can manually classify. Finding potentially habitable worlds requires analyzing complex signals buried in noisy data. I wanted to build an accessible, highly accurate, and fully transparent machine learning application that allows anyone to classify Kepler Objects of Interest (KOI) in real time.

🚀 What it does

Kepler Exoplanet Oracle is an end-to-end classification engine and mission-control styled dashboard. Users input key telescope measurements—such as Orbital Period (koi_period), Planetary Radius (koi_prad), and Stellar Effective Temperature (koi_steff)—and instantly receive an AI-driven prediction classifying the candidate object as CONFIRMED, CANDIDATE, or FALSE POSITIVE.

🛠️ How I built it

Data Processing & EDA: Preprocessed the KOI_Cumulative_clean.csv dataset. To resolve severe class imbalance between false positives and confirmed exoplanets, I applied SMOTE (Synthetic Minority Over-sampling Technique). Missing attributes were imputed using median statistical strategies.
Domain-Aware Feature Selection: Isolated physical telemetry variables—orbital timing, host star temperature, and planet radius relative to Earth—to train domain-accurate decision boundaries.
Machine Learning Pipeline: Trained a Random Forest Classifier using Scikit-Learn, achieving an overall 95% accuracy on unseen test data.
Explainable AI (XAI): Integrated SHAP (SHapley Additive exPlanations) to generate feature importance summaries, converting black-box model decisions into clear, verifiable insights.
Backend API: Built a high-performance REST API with FastAPI and Uvicorn to handle request payloads and model inference via joblib.
Frontend UI: Designed a custom sci-fi themed dashboard using HTML5, CSS3, and JavaScript, complete with a visual flux monitor and live status output.

⚡ Challenges I ran into

Class Imbalance: Raw telescope telemetry naturally skews heavily toward false positives. Applying SMOTE without overfitting required careful validation tuning.
API Pipeline Preprocessing: Handled live payload edge cases in FastAPI by enforcing baseline imputation routines for non-input parameters.
Clean Repository Architecture: Properly configuring .gitignore to exclude large Python virtual environments while keeping essential serialized .pkl models tracked.

🎉 Accomplishments that I'm proud of

Developed and deployed the entire end-to-end full-stack machine learning application as a solo developer.
Reached 95% classification accuracy while maintaining full model interpretability through SHAP.
Built a custom, responsive, sci-fi UI that transforms raw data prediction into an engaging user experience.

🎓 What I learned

Advanced multi-class resampling strategies using SMOTE.
Productionizing machine learning models with FastAPI and Uvicorn.
Communicating complex AI predictions using SHAP feature attribution.

🔮 What's next for Kepler Exoplanet Oracle

Deploying the FastAPI backend and frontend to cloud hosts like Render or Vercel.
Adding real-time transit light-curve visualizations directly into the UI.