A complete bilingual product site goes live
English becomes the default public edition, with matching English and Chinese product sections, update history, community entry points, structured data, and feedback flows.
BUILDING IN PUBLIC · SINCE JULY 2026
PRISMS AI update history
A crawlable record of architecture, product, knowledge, and live-runtime work.
Put the AI hub back on the student's iPhone
PRISMS AI defines the boundary between on-device agents, the stateless model relay, the isolated control plane, provenance-first memory, and consented iOS system capabilities.
A clearer local-first student workspace
Ask, memory, personal knowledge, agent stages, and task-appropriate text, image, audio, or interactive HTML deliverables move into one persistent iOS workspace.
Give every answer a clearer source boundary
Business agents retrieve school and external knowledge before source-dependent answers. Unsupported claims remain stable knowledge gaps instead of model guesses.
The student experience becomes iOS-first
The private student assistant moves to iOS while the public website focuses on product architecture, dated updates, feedback, and community.
Calibrate the agent runtime with real failures
End-to-end student tests begin treating the actual agent chain, evidence consumption, and final artifact—not a plausible Intake response—as completion criteria.
Move campus knowledge from static files to auditable sources
A full refresh checks 517 URLs, saves 502, and keeps web pages, PDFs, documents, and access errors in distinct inspectable states.
Expand official PRISMS knowledge coverage
Official school ingestion grows to 517 URLs and 592 document pages, updating 286 knowledge objects and 2,407 evidence records.
Replace blind retries with reasoned correction
Agents select a correction skill from the actual stage and failure type, preserve valid work, and rebuild the complete failed object coherently.
Drive Campus OS with multi-domain student work
Continuous scenarios cover courses, dining, clubs, residence life, communication, and campus services with real-source and human-confirmation requirements.
Build a traceable foundation for campus facts
The first expanded school-site scan fetches 160 URLs, saves 151, and updates 143 knowledge objects with 932 evidence records.
The 2026–27 course catalog enters the knowledge base
Fifty-eight courses are rebuilt from the official catalog with department, credit, level, recommended grade, offering status, and page-level evidence.