Skip to content
pmLet’s talk
All projects

GRIT-X-AWA

Exoplanet classification over NASA Kepler and TESS data

GRIT-X-AWA is a web app for classifying exoplanet candidates from two NASA missions, Kepler and TESS. The front end is Astro 5 with React 18 islands and a three.js starfield: a landing page drops into a Mission Control dashboard where you pick a mission, upload a CSV or type a candidate in by hand, read model-metric cards, then open a full-page results modal with per-class confidence and CSV or JSON export. A paginated browser reads Supabase-hosted mission datasets, react-three-fiber renders 3D orbital views, and an in-app chatbot answers questions about the science.

GRIT-X-AWA: project preview
Role
NASA Space Apps team · full-stack + ML
Year
2026
Platform
Software project
Status
prototype

The problem

Vetting a transit signal as a real planet, an eclipsing binary, or noise is slow expert work, and the raw Kepler and TESS tables are wide and unfriendly. GRIT-X-AWA puts a trained ensemble behind a single upload box so a candidate list becomes labeled dispositions with confidence in seconds, and makes the pipeline explorable in the browser instead of a notebook.

What I built

The work includes frontend, backend, kepler + tess models.

The choices behind it.

NASA Space Apps team · full-stack + ML

Implementation choice

Two per-mission ensembles of three gradient-boosting models each (CatBoost, XGBoost, LightGBM) combine by a fixed weighted soft vote, argmax(0.40*Cat + 0.35*XGB + 0.25*LGBM), with the weights and class order pinned in each model's meta.json.

Implementation choice

Kepler reads 21 KOI features into 3 classes (CANDIDATE, CONFIRMED, FALSE POSITIVE); TESS reads 17 base features engineered into about 66 model inputs across 6 classes (APC, CP, FA, FP, KP, PC), and the backend auto-detects which mission a CSV belongs to.

Implementation choice

A keyless demo mode is on by default: CSV upload, manual entry, the Supabase data browser, and the chatbot all fall back to deterministic canned fixtures and bundled sample rows, so the deployed site never shows a blank screen or asks for a key; PUBLIC_DEMO=false restores the real backend paths.

Implementation choice

The front end runs Astro 5 in SSR mode on the Vercel adapter with React 18 islands: a three.js starfield landing (SpaceScene), a Mission Control dashboard, a react-three-fiber 3D orbital viewer (ExoplanetVisualization3D), and a full-page prediction-results modal that exports to CSV and JSON via html2canvas.

Implementation choice

The AI chatbot is an Astro server endpoint (/api/chat) with per-IP rate limiting (20 messages per day) that degrades to a scripted reply when no model key is present, so the assistant stays graceful in the public demo.

Implementation choice

The FastAPI service loads and caches the pickled models on startup and exposes CSV upload, single predict, stats, and recent-prediction endpoints; Supabase (Postgres plus Storage) persists uploads and prediction history when configured, and the service still runs in a local mode when the database is absent.

gradient-boosting models: a CatBoost, XGBoost, LightGBM ensemble per mission, blended 0.40 / 0.35 / 0.25
6
NASA missions, 9 disposition classes: Kepler's 3 and TESS's 6
2
keys or backend needed: demo mode is on by default with canned predictions and bundled sample data
0
Kepler features and 17 TESS features (engineered into about 66) feed the models
21

Under the surface.

frontend

app

Astro 5 and React 18 client with a three.js starfield: the Mission Control dashboard (mission pick, CSV upload or manual entry, metric cards, results modal with CSV and JSON export), a paginated Supabase data browser, react-three-fiber orbital views, and the chatbot.

backend

service

FastAPI app on Google Cloud Run that auto-detects the mission, preprocesses the CSV, runs the weighted ensemble, and returns per-row predictions with per-class confidence over a REST API.

kepler + tess models

model

Two per-mission ensembles (CatBoost, XGBoost, LightGBM) plus their imputer, encoders, and meta.json, pickled and loaded on backend startup for the 3-class Kepler and 6-class TESS vote.

demoFixtures

library

The keyless demo layer: deterministic canned prediction fixtures, bundled sample rows for the data browser, and a scripted chatbot reply, gated by the PUBLIC_DEMO flag so the site runs with no backend or keys.

chat endpoint

service

An Astro SSR /api/chat route with per-IP daily rate limiting that powers the in-app assistant and degrades to a graceful scripted reply in demo mode.

Build details Stack, scope, and implementation

Built with

  • Astro 5
  • React 18
  • TypeScript 5
  • three.js r167
  • react-three-fiber
  • @react-three/drei
  • Tailwind CSS v3
  • Supabase
  • Python
  • FastAPI
  • XGBoost
  • LightGBM
  • CatBoost
  • scikit-learn
  • pandas
  • Google Cloud Run
  • Vercel

Areas

  • exoplanet classification
  • ensemble ML
  • NASA open data
  • three.js
  • FastAPI
  • data visualization

More about the project

The prediction engine is a Python FastAPI service on Google Cloud Run. Each mission has its own ensemble of three gradient-boosting models, CatBoost, XGBoost, and LightGBM, blended by a fixed weighted soft vote (0.40 / 0.35 / 0.25) into a labeled disposition. It now ships as a keyless public demo: demo mode is on by default, so uploads and manual entries return canned but realistic predictions, the data browser falls back to bundled sample rows, and the chatbot returns a graceful scripted reply, which means the whole site runs with no backend and no API keys. Set PUBLIC_DEMO=false with a live backend to hit the real ensemble.

Try the product.

Interactive demo · Open full screen