The problem
Screenshots, files, scoring rules, and feedback often live in separate places, making it difficult to revisit why a model received a particular assessment.
My contribution
- Led workflow and report requirements; used AI-assisted implementation.
- Reviewed how evidence, human revisions, and access rules fit together.
Working approach: product-led, AI-assisted implementation, followed by review and iteration.
Decisions & tradeoffs
01
Keep a revision trail
- The choice
- Create report versions and preserve their relationship to the inputs and previous revision.
- The tradeoff
- A generated assessment remains a draft that needs human review.
- How it is checked
- The repository documents version snapshots and includes tests for core report and permission logic.
02
Share a limited view
- The choice
- Offer explicit read-only sharing while keeping internal report metadata out of the public projection.
- The tradeoff
- The full workspace requires invited access and a configured model provider.
- How it is checked
- A shared access layer and field allowlist define the public view.
What the work shows
The public code connects task setup, screenshot and artifact review, versioned reports, and export in one application.
Follow the evidence
What this does not establish
- The hosted workspace uses invited access; its link is not an anonymous full-feature demo.
- AI-written comparisons do not constitute independent benchmark certification.