AI tooling for scientific model metadata
Built full-stack tooling for managing machine-learning model metadata and developed an AI-assisted workflow for extracting structured information from scientific PDFs, with validation and GitHub-based automation around the wider open-science workflow.
- Context
- Scientific model metadata tooling
- Organization
- Ersilia
- Role
- Fullstack Developer
- Period
- 2024–2025
Problem
Machine-learning model metadata needed a structured application and backend rather than a collection of disconnected manual steps.
Relevant information in scientific PDFs had to be extracted into a form that could be reviewed and used by the wider open-science workflow.
Constraints
AI-extracted information could not be treated as automatically correct.
The application had to fit an existing GitHub-based collaboration and repository workflow.
My responsibility
Built the full-stack Next.js and TypeScript application.
Designed the database schema and backend for structured model metadata.
Developed the scientific-PDF workflow with the Vercel AI SDK.
Automated metadata validation and repository synchronization with GitHub API and Actions.
Worked directly with researchers and engineers in the open-science team.
Approach
Modeled metadata explicitly so the application, validation, and repository workflow shared the same vocabulary.
Kept AI extraction inside a reviewable workflow instead of presenting model output as a final source of truth.
Used GitHub automation to connect application work with the repositories already used by the team.
Key technical decisions
Treat AI output as input to validation
Scientific PDFs vary in structure and LLM output is probabilistic.
Extract structured candidates through the AI workflow, then keep validation and human review explicit.
Review adds friction, but avoids making unsupported metadata authoritative.
Integrate with the existing repository workflow
The open-science team already used GitHub repositories as part of its operating model.
Automate validation and synchronization through the GitHub API and GitHub Actions instead of creating a separate publishing path.
Outcomes
Delivered a full-stack interface and backend for structured model metadata.
Connected AI-assisted PDF analysis to explicit validation.
Reduced disconnects between application data and GitHub-based open-science workflows through automation.
Reflection
AI features become dependable when the surrounding data model, validation, and ownership are designed as carefully as the model call itself.
Related work
Need someone who can own the boundaries?
I am open to selected product work and future senior engineering opportunities.
Contact me