As offensive tooling becomes increasingly autonomous, the line between detection and prevention keeps moving. My current focus is building systems that learn the intent behind an attack rather than the signature.
The Challenge
NASA's TESS mission generates terabytes of light curve data. Traditional ML models can classify transit signals with high accuracy, but astronomers need to understand why a prediction was made before investing telescope time.
Our Approach: Explainability First
Instead of treating the model as a black box, we architected North-Star around interpretability:
1. Feature Engineering
We extracted physically meaningful features:
- Transit depth and duration
- Periodicity and phase coherence
- Stellar noise characteristics
2. SHAP Values
For each prediction, we compute SHAP (SHapley Additive exPlanations) scores that quantify each feature's contribution.
3. Visual Explanations
The UI overlays model attention maps on light curves, highlighting exactly which time windows influenced the decision.
Validation
We tested on confirmed exoplanets from Kepler archives:
- 96.4% precision
- 93.1% recall
- 100% of false positives traced to known stellar activity (flares, starspots)
Impact
Astronomers using North-Star reported 40% faster triage compared to manual review. The explainability layer caught three edge cases where high-confidence predictions were actually instrumental artifacts—saving thousands of dollars in follow-up observations.
Open Science
We open-sourced the entire pipeline on GitHub. The astronomy community has since extended it to detect binary star systems and stellar oscillations.