Live research demo | Research trace: data -> protocol -> results
SlopeSense is an experimental landslide research application. It is not an official warning service and is not validated for safety decisions. Real risk scoring returns 503 until the independent data and operational validation gates clear.
- Six representative locations receive real Open-Meteo provider snapshots on an hourly retrieval plan. These are modeled weather/soil data, not local gauges, NASA satellite inputs or a validated regional landslide forecast.
- API keys and contacts use persistent storage. Alert delivery is disabled even when credentials are present. CAP examples are Test, not government warnings.
- Reports prominently label sample material. Retrospective validation reports not_validated rather than fabricated success or confidence intervals.
- Three research stages retain real datasets, frozen protocols, failures and fitted exploratory models. Stage3 uses the official current NASA COOLR catalog.
Read every stage and download its artifacts. None passed the 90%recall/10%matched-background false-positive research screen. Stage3 unseen2017+ confirmations caught27.9%temporal and23.5%spatial reported events at the fixed0.5 threshold. They are case-control classification results, not population probabilities.
The old1km/FPI,24-48hour warning, calibrated95%confidence, NDMA compatibility, retrospective success and alert-delivery performance claims are not established. There is no validated warning lead time, field-tested coverage or zero-downtime guarantee.
Start with Stage1 sources/protocol, Stage2 protocol and Stage3 protocol. Results: Stage1, Stage2, Stage3. Source licenses and archive paths are in the research index. NASA provenance does not prove event precision, monitored non-events/base rates, pre-event terrain or prospective operational performance.
Pretrained-model candidates and limitations. External model weights remain their authors' work, not SlopeSense's model.
Backend: FastAPI/Python. Frontend: Next.js. Database: PostgreSQL. Review
backend/config.py, docker-compose.yml, requirements.txt and frontend/package.json
for configuration. Keep credentials out of the repository. Do not add payment cards
or paid services for this noncommercial research deployment.
pip install -r requirements.txt
pytest
uvicorn backend.api.main:app --reload
cd frontend
npm install
npm run devLegacy synthetic trainers are not the current research route. Offline fitted artifacts under training/ are not wired to production warning scoring.
Project code: Apache2.0. Source data and third-party models retain their own licenses; see source notices. This project does not claim NASA or GSI endorsement. NASA LHASA uses NOSAv1.3, not the project's Apache license.