We design an offline generative AI pipeline that turns course materials and assessment evidence into reusable, personalised remediation videos.
EgoBlind-RA routes egocentric queries from blind and low-vision users by urgency using a lightweight CLIP classifier (0.905 ROC-AUC), then allocates response policies matched to their safety stakes — showing that the value of DPO and the cost of routing noise both depend sharply on architectural regime.
We present StarCLIP, a contrastive learning framework aligning APOGEE and JWST/NIRSpec spectra into a unified latent space, enabling accurate stellar property recovery and demonstrating the feasibility of foundation models for future spectroscopic surveys.
We analyzed light curves of 16 PRGs in omega Centauri, confirming the prevalence of LSPs and establishing period–luminosity sequences consistent with Population II red giants elsewhere.