FIELD NOTES / Scientific AI Systems
AI4Science & Scientific Agents
How can scientific tasks combine tools, explicit constraints, candidate generation and evaluation?
Last reviewed 2026-10-19 · Non-sensitive overview of private work
Overview
Computational scientific discovery · AI4Science & Scientific Agents
参与科学任务工作流的团队开发，研究工具调用、候选方案评估与任务约束。
Research Question
How can scientific tasks combine tools, explicit constraints, candidate generation and evaluation?
Technical Approach
Team repositories implement screening and route-first molecular candidate workflows, a staged planner/tool/evaluator loop, optional LLM narration, rule-based fallback, containerized execution and scoring checks.
My Contribution
Participated in the team project. Personally attributable LeadForge commits cover constraint/risk architecture, product and scientific communication diagrams, and diagram rendering workflow. Individual implementation ownership of the molecular engines is not established by the inspected history.
Evidence
- Executable team pipeline and knowledge-loader tests inspected.
- Nine attributable LeadForge commits reviewed; predominantly architecture/communication figures.
- Competition archive and local development packages located.
Outputs
- Team scientific task workflows
- Architecture and risk-control communication artifacts
Current Status
Competition ranking, biological efficacy and any advantage from LLM use remain unverified. The implementation is a team effort; my documented contributions are described above.
Related Publications
No project-specific peer-reviewed publication is listed. Ongoing work is kept separate from the published bibliography.