GenomeCanvas
A 3D protein explorer that shows structure confidence and lets typed UI commands guide the scene before optional model narration begins.
Problem
Exploring protein structures and their linked disease or drug records involves disconnected views. An interface should help users navigate the underlying data while keeping its limits visible.
Approach
Built a Three.js scene, Mol* structure view and graph workspace over local fixtures. Grounded answers and resolved entity commands are constructed first; an optional Claude layer rephrases the answer, with a local fallback when narration is unavailable.
Impact
The recorded fixture contains 54 proteins, with 52 real AlphaFold structures and two fallbacks. Similarity means visual proximity in reduced UMAP space, not biological equivalence. Deployment configuration exists, but this source snapshot does not establish a public deployment.
Key Metrics
Technologies
Links
My Role
Defined the interaction model and specification and directed AI-assisted implementation, including grounded commands, visualization behavior and correctness checks.