Computational Software Development

Build production-oriented biomedical software with a bias toward reproducibility, clear interfaces, and testable workflows.

Workflow

Clarify the product boundary: analysis script, library, API, web app, pipeline, or multi-agent system.
Inspect the local repository before proposing architecture changes; preserve existing conventions unless they are actively harmful.
Choose the simplest stack that fits the job: plain Python package first, then API or UI layers only when they solve a real need.
Separate scientific logic, orchestration, and presentation so domain code remains testable outside the UI.
Add validation around files, schemas, model outputs, and external API calls; fail loudly when assumptions break.
Prefer deterministic scripts for repeated scientific tasks and reserve LLM calls for synthesis, extraction, ranking, or planning.
Deliver runnable code plus concise usage notes, tests, and environment assumptions.

Keep domain logic in importable modules rather than notebooks or route handlers.
Treat data contracts explicitly: define expected columns, shapes, units, genome builds, and identifier namespaces.
For LLM features, specify provider fallback rules, timeout budgets, retry policy, and citation or provenance requirements.
For biomedical applications, avoid hidden PHI flows and minimize access to sensitive fields.
Prefer small composable services over one large agent script.

Python packages with src/ layout, CLI entry points, and tests.
Flask or FastAPI backends for analysis services or RAG endpoints.
Data pipelines for omics preprocessing, QC, and result packaging.
Multi-LLM orchestration with structured outputs, adjudication, and evaluation.
Research dashboards that expose results without burying the computational workflow.

Read references/architecture-patterns.md for architecture selection and delivery checklists.
Read references/biomedical-llm-patterns.md when the request involves RAG, citations, or multi-provider systems.