ScheduleSolve AI

All code was written during the hackathon (Oct 27 - Nov 4, 2026); no pre-existing product, the first commit is the project skeleton. IMPACT. Every term a coordinator rebuilds the course timetable by hand, reconciling constraints, degree roadmaps, rooms and a campus you cannot cross in ten minutes. It takes days and is stale the moment a professor leaves. I have watched this happen every term from inside a university: Build for Your World. Universities do not lack software; they lack proof. AACRAO's 2025 survey of 471 institutions: "Our entire schedule is built manually, one course at a time." Who it changes: the coordinator gets an afternoon instead of a week; every professor a timetable that provably honours their hard constraints; every student a plan re-verified against the timetable, or an honest refusal when the fitting section does not exist. It could ship: one command brings up Postgres, API and UI; the model persists across terms; the registrar sets load rules, track factors and the credit band as policy, not constants; and it runs on Meta's open weights by configuration. Validated on real data: ingestion ran against Rutgers-Camden's public 2026-27 catalog (576 pages) at ~640 tokens/page, naming missing inputs rather than inventing them. TECHNICAL DEPTH. The model never writes the schedule. Muse Spark turns catalogs, degree maps, room sheets and scans into one typed model, and a sentence into validated operations; two independent solvers - Google's CP-SAT and our own C++17 engine - prove every placement and repair before it is shown. An independent verifier re-derives every hard constraint from the placements alone. Both return the same proven optimum; ours is ~4x faster at 41 sections (35 vs 137 ms) and the only one that finishes a 361-section, 12,530-student institution in ~5.6 s, where CP-SAT reports unknown. Impossible is an answer: the solver isolates the minimal conflict core (either engine; a test proves they agree), and every repair is re-solved before it is shown. 183 tests, CI on every push, a benchmark that regenerates every number, a 15-scenario audit with zero errors. META TOOLING. The surprise: the model proposes institutional moves - raise a load cap, split a section, cross-train, open an elective - and deterministic math says which are real, because every one is re-solved before it is shown. It is ambitious because it is never the authority. Roles run on Muse Spark Open on Baseten (18/18 live calls); with no key every role degrades to a labelled deterministic fallback. DEMO. Live at https://prohle--schedulesolve-ai-ui-web.us-east.modal.run (sample sign-in, no API key) with a 2:59 video recorded live against Baseten. Every beat is clickable: import, intake, a dual-engine race ranked from outputs alone, term change with zero-ripple cover, a conflict core with ranked repairs, and a student's plan. Honest boundaries: single-tenant, demo-grade auth, no registrar has used it in production.