Full-Stack Scientific Software Engineering
From FastAPI to PyTorch Bindings and Climate Modeling
2026-06-13
Overview
- Motivation
- Practice language listening skills
- Small segment of YouTube video
- Your transcript vs reference
- Technology
- GitHub: lingua-loop
- Python (FastAPI)
- CI/CD (GitHub Actions)
- Automated Testing (Pytest)
- Public Documentation
- PyPI: lingua-loop
docker run -p 49152:49152 lingua-loopSystem Architecture
- API/Web Layer
- Handle client HTTP requests
- Handle web server responses
- Business Logic/Services Layer
- Parse and normalize user input
- Compute scoring metrics for transcription attempt
- Data Layer
- Use “boundary” (aka integrations) to scrape reference YouTube transcripts
- Read/write to SQLite database
Testing Strategy
tests/
├── boundary/
│ └── test_youtube_transcript_api.py
├── integration/
│ └── test_integration.py
└── unit/
├── api/
├── db/
└── services/- Multi-layer testing
- Extensive use of mocking to isolate dependencies (DB, services, external APIs)
- Unit tests validate service logic in complete isolation
- Integration tests run full stack: API → service → database (no mocks)
- CI/CD on GitHub enforces all tests pass during PR
From Code to Package
Overview
- Library for coupling PyTorch ML models directly to Fortran code
- Used in Community Earth System Model, ICON Model, and UK Atomic Energy Authority
- Developed primarily by University of Cambridge
Overview
- Unified global weather prediction and climate model co-developed by German Weather Service, MeteoSwiss, etc.
- Massively parallel with MPI, OpenMP, and OpenACC offloading
- Operational at DWD since 2015 for weather forecasting in Germany and Switzerland